diff --git a/.gitignore b/.gitignore index 7fa2022..0f5657d 100644 --- a/.gitignore +++ b/.gitignore @@ -2,4 +2,12 @@ .idea venv .venv -*.db \ No newline at end of file +*.dbsrc/inputs/*.pdf +src/outputs/*.pdf +src/inputs/*.pdf +src/outputs/*.pdf +fireform.db +*.bak +ngrok.exe +out.txt +benchmark_proof.py diff --git a/api/db/models.py b/api/db/models.py index f76c93b..c8c166d 100644 --- a/api/db/models.py +++ b/api/db/models.py @@ -15,4 +15,28 @@ class FormSubmission(SQLModel, table=True): template_id: int input_text: str output_pdf_path: str - created_at: datetime = Field(default_factory=datetime.utcnow) \ No newline at end of file + created_at: datetime = Field(default_factory=datetime.utcnow) + +# ADD THIS TO api/db/models.py +# (append to existing file — don't replace) + +from sqlmodel import SQLModel, Field +from typing import Optional +from datetime import datetime + + +class IncidentMasterData(SQLModel, table=True): + """ + The Incident Data Lake. + Stores all extracted data from one incident as a master JSON blob. + Any agency can generate their PDF from this single record — zero new LLM calls. + """ + id: Optional[int] = Field(default=None, primary_key=True) + incident_id: str = Field(index=True) # INC-2026-0321-4821 + master_json: str # JSON string — all extracted fields + transcript_text: str # original transcript + location_lat: Optional[float] = None # from PWA GPS + location_lng: Optional[float] = None # from PWA GPS + officer_notes: Optional[str] = None # additional context + created_at: datetime = Field(default_factory=datetime.utcnow) + updated_at: datetime = Field(default_factory=datetime.utcnow) \ No newline at end of file diff --git a/api/db/repositories.py b/api/db/repositories.py index 6608718..11f6549 100644 --- a/api/db/repositories.py +++ b/api/db/repositories.py @@ -1,19 +1,117 @@ from sqlmodel import Session, select from api.db.models import Template, FormSubmission -# Templates + +# ── Templates ───────────────────────────────────────────────── + def create_template(session: Session, template: Template) -> Template: session.add(template) session.commit() session.refresh(template) return template + def get_template(session: Session, template_id: int) -> Template | None: return session.get(Template, template_id) -# Forms + +def get_all_templates(session: Session, limit: int = 100, offset: int = 0) -> list[Template]: + statement = select(Template).offset(offset).limit(limit) + return session.exec(statement).all() + + +# ── Forms ───────────────────────────────────────────────────── + def create_form(session: Session, form: FormSubmission) -> FormSubmission: session.add(form) session.commit() session.refresh(form) - return form \ No newline at end of file + return form + + +def get_form(session: Session, submission_id: int) -> FormSubmission | None: + return session.get(FormSubmission, submission_id) + + +# ADD THESE FUNCTIONS TO api/db/repositories.py +# (append to existing file — don't replace) + +import json +from api.db.models import IncidentMasterData +from datetime import datetime + + +def create_incident(db, incident: IncidentMasterData) -> IncidentMasterData: + db.add(incident) + db.commit() + db.refresh(incident) + return incident + + +def get_incident(db, incident_id: str) -> IncidentMasterData: + from sqlmodel import select + return db.exec( + select(IncidentMasterData).where( + IncidentMasterData.incident_id == incident_id + ) + ).first() + + +def get_all_incidents(db) -> list: + from sqlmodel import select + return db.exec(select(IncidentMasterData)).all() + + +def update_incident_json(db, incident_id: str, new_data: dict, new_transcript: str = None) -> IncidentMasterData: + """ + Smart Merge new extracted data into existing master JSON to enable + Collaborative Incident Consensus. Protects existing data from being + wiped by LLM `null` hallucinations, and appends long-form text. + """ + incident = get_incident(db, incident_id) + if not incident: + return None + + existing = json.loads(incident.master_json) + + for key, value in new_data.items(): + # 1. Ignore empty/null values to protect existing data + if value is None or str(value).strip().lower() in ("null", "none", "", "n/a"): + continue + + # 2. If the field exists, handle smart merging vs overwriting + if key in existing and existing[key]: + old_value = existing[key] + + # Use string representation for safe comparison + old_str = str(old_value).strip() if not isinstance(old_value, list) else "\n".join(str(i) for i in old_value) + new_str = str(value).strip() if not isinstance(value, list) else "\n".join(str(i) for i in value) + + # If the value is identical, do nothing + if old_str.lower() == new_str.lower(): + continue + + # If it's a long-form text field (Notes, Description, Narrative, Summary, etc) + long_fields = ("note", "desc", "narrative", "summary", "remark", "detail", "comment") + if any(lf in key.lower() for lf in long_fields): + # Prevent recursive appending + if new_str not in old_str: + existing[key] = f"{old_str}\n\n[UPDATE]: {new_str}" + else: + # Standard Field Correction (e.g. ID, City) - overwrite the old value + existing[key] = value + else: + # 3. Brand new field + existing[key] = value + + incident.master_json = json.dumps(existing) + + # Safely append the new transcript segment for true consensus history + if new_transcript and new_transcript.strip() not in incident.transcript_text: + incident.transcript_text = f"{incident.transcript_text}\n\n---\n[UPDATE]: {new_transcript.strip()}" + + incident.updated_at = datetime.utcnow() + db.add(incident) + db.commit() + db.refresh(incident) + return incident \ No newline at end of file diff --git a/api/main.py b/api/main.py index d0b8c79..9b2cd98 100644 --- a/api/main.py +++ b/api/main.py @@ -1,7 +1,32 @@ -from fastapi import FastAPI -from api.routes import templates, forms +from fastapi import FastAPI, Request +from fastapi.middleware.cors import CORSMiddleware +from fastapi.responses import JSONResponse +from fastapi.staticfiles import StaticFiles +from api.routes import templates, forms, transcribe, incidents +from api.errors.base import AppError +from typing import Union +import os app = FastAPI() +app.add_middleware( + CORSMiddleware, + allow_origins=["*"], + allow_methods=["*"], + allow_headers=["*"], +) + +@app.exception_handler(AppError) +def app_error_handler(request: Request, exc: AppError): + return JSONResponse( + status_code=exc.status_code, + content={"detail": exc.message} + ) + app.include_router(templates.router) -app.include_router(forms.router) \ No newline at end of file +app.include_router(forms.router) +app.include_router(transcribe.router) +app.include_router(incidents.router) + +if os.path.exists("mobile"): + app.mount("/mobile", StaticFiles(directory="mobile", html=True), name="mobile") \ No newline at end of file diff --git a/api/routes/forms.py b/api/routes/forms.py index f3430ed..9df6a5c 100644 --- a/api/routes/forms.py +++ b/api/routes/forms.py @@ -1,25 +1,140 @@ +import os from fastapi import APIRouter, Depends +from fastapi.responses import FileResponse from sqlmodel import Session from api.deps import get_db -from api.schemas.forms import FormFill, FormFillResponse -from api.db.repositories import create_form, get_template +from api.schemas.forms import FormFill, FormFillResponse, BatchFormFill, BatchFormFillResponse, BatchResultItem +from api.db.repositories import create_form, get_template, get_form from api.db.models import FormSubmission from api.errors.base import AppError from src.controller import Controller +from src.llm import LLM +from src.filler import Filler router = APIRouter(prefix="/forms", tags=["forms"]) + @router.post("/fill", response_model=FormFillResponse) -def fill_form(form: FormFill, db: Session = Depends(get_db)): - if not get_template(db, form.template_id): +async def fill_form(form: FormFill, db: Session = Depends(get_db)): + template = get_template(db, form.template_id) + if not template: raise AppError("Template not found", status_code=404) - fetched_template = get_template(db, form.template_id) + if not os.path.exists(template.pdf_path): + raise AppError(f"Template PDF not found: {template.pdf_path}", status_code=404) + + try: + # Step 1: LLM Extraction (Async) + llm = LLM(transcript_text=form.input_text, target_fields=template.fields) + await llm.async_main_loop() + extracted_data = llm.get_data() - controller = Controller() - path = controller.fill_form(user_input=form.input_text, fields=fetched_template.fields, pdf_form_path=fetched_template.pdf_path) + # Step 2: PDF Filling (Sync) + # Using filler directly to avoid redundant extraction in controller + filler = Filler() + path = filler.fill_form_with_data( + pdf_form=template.pdf_path, + data=extracted_data + ) + except Exception as e: + raise AppError(f"Processing failed: {str(e)}", status_code=500) + + if not path or not os.path.exists(path): + raise AppError("PDF generation failed.", status_code=500) submission = FormSubmission(**form.model_dump(), output_pdf_path=path) return create_form(db, submission) +@router.post("/fill/batch", response_model=BatchFormFillResponse) +async def fill_batch(batch: BatchFormFill, db: Session = Depends(get_db)): + if not batch.template_ids: + raise AppError("template_ids must not be empty", status_code=400) + + templates = [] + for tid in batch.template_ids: + tpl = get_template(db, tid) + if not tpl or not os.path.exists(tpl.pdf_path): + raise AppError(f"Template {tid} invalid or PDF missing", status_code=404) + templates.append(tpl) + + # Step 1: LLM Extraction (Async - ONE call for all templates) + merged_fields = {} + for tpl in templates: + if isinstance(tpl.fields, dict): merged_fields.update(tpl.fields) + else: + for f in tpl.fields: merged_fields[f] = f + + try: + llm = LLM(transcript_text=batch.input_text, target_fields=merged_fields) + await llm.async_main_loop() + extracted_json = llm.get_data() + except Exception as e: + raise AppError(f"Extraction failed: {str(e)}", status_code=500) + + # Step 2: PDF Filling (Sync - per template) + results = [] + success_count = 0 + filler = Filler() + + for tpl in templates: + try: + tpl_field_keys = list(tpl.fields.keys()) if isinstance(tpl.fields, dict) else tpl.fields + tpl_data = {k: extracted_json.get(k) for k in tpl_field_keys} + + output_path = filler.fill_form_with_data(pdf_form=tpl.pdf_path, data=tpl_data) + + submission = FormSubmission( + template_id=tpl.id, + input_text=batch.input_text, + output_pdf_path=output_path + ) + saved = create_form(db, submission) + + results.append(BatchResultItem( + template_id=tpl.id, + template_name=tpl.name, + success=True, + submission_id=saved.id, + download_url=f"/forms/download/{saved.id}" + )) + success_count += 1 + except Exception as e: + results.append(BatchResultItem( + template_id=tpl.id, + template_name=tpl.name, + success=False, + error=str(e) + )) + + return BatchFormFillResponse( + total=len(templates), + succeeded=success_count, + failed=len(templates)-success_count, + results=results + ) + + +@router.get("/{submission_id}", response_model=FormFillResponse) +def get_submission(submission_id: int, db: Session = Depends(get_db)): + submission = get_form(db, submission_id) + if not submission: + raise AppError("Submission not found", status_code=404) + return submission + + +@router.get("/download/{submission_id}") +def download_filled_pdf(submission_id: int, db: Session = Depends(get_db)): + submission = get_form(db, submission_id) + if not submission: + raise AppError("Submission not found", status_code=404) + + file_path = submission.output_pdf_path + if not os.path.exists(file_path): + raise AppError("PDF file not found on server", status_code=404) + + return FileResponse( + path=file_path, + media_type="application/pdf", + filename=os.path.basename(file_path) + ) \ No newline at end of file diff --git a/api/routes/incidents.py b/api/routes/incidents.py new file mode 100644 index 0000000..a63171d --- /dev/null +++ b/api/routes/incidents.py @@ -0,0 +1,268 @@ +import os +import json +from fastapi import APIRouter, Depends +from fastapi.responses import FileResponse +from sqlmodel import Session +from api.deps import get_db +from api.db.models import IncidentMasterData, FormSubmission +from api.db.repositories import ( + create_incident, get_incident, get_all_incidents, + update_incident_json, get_template, create_form +) +from api.errors.base import AppError +from src.filler import Filler +from src.llm import LLM +from src.controller import Controller +from datetime import datetime + +router = APIRouter(prefix="/incidents", tags=["incidents"]) + + +# ── Schema: Extract & Store ────────────────────────────── + +@router.post("/extract") +async def extract_to_data_lake( + input_text: str, + incident_id: str = None, + location_lat: float = None, + location_lng: float = None, + db: Session = Depends(get_db) +): + """ + THE CORE DATA LAKE ENDPOINT. + + Extracts ALL possible fields from transcript and stores as + Master Incident JSON. No template needed — extracts everything. + Later: any agency generates their PDF from this stored data + without re-running the LLM. + + If incident_id already exists — merges new data into existing. + This supports multi-officer reports: each officer adds their + perspective, system merges into one master record. + """ + if not incident_id: + # Auto-generate incident ID + now = datetime.utcnow() + incident_id = f"INC-{now.year}-{now.month:02d}{now.day:02d}-{now.hour:02d}{now.minute:02d}" + + print(f"[DATA LAKE] Extracting incident: {incident_id}") + + # Get all templates to build maximum superset of fields + from api.db.repositories import get_all_templates + all_templates = get_all_templates(db) + + # Start with an empty schema to allow fully dynamic LLM extraction + # The LLM will use any uploaded template fields as a base guide, + # and autonomously invent new fields for the rest. + merged_fields = {} + + if all_templates: + # Build superset from all known templates + for tpl in all_templates: + if isinstance(tpl.fields, dict): + merged_fields.update(tpl.fields) + print(f"[DATA LAKE] Base schema: {len(merged_fields)} template fields across {len(all_templates)} templates") + + try: + llm = LLM(transcript_text=input_text, target_fields=merged_fields) + await llm.async_main_loop() + extracted = llm.get_data() + print(f"[DATA LAKE] Extracted {len(extracted)} fields") + except ConnectionError: + raise AppError("Could not connect to Ollama.", status_code=503) + except Exception as e: + raise AppError(f"Extraction failed: {str(e)}", status_code=500) + + # Check if incident already exists — merge if so + existing = get_incident(db, incident_id) + if existing: + print(f"[DATA LAKE] Merging into existing incident {incident_id}") + updated = update_incident_json(db, incident_id, extracted, new_transcript=input_text) + return { + "incident_id": incident_id, + "status": "merged", + "fields_extracted": len(extracted), + "total_fields": len(json.loads(updated.master_json)), + "message": f"Merged into existing incident. Total fields: {len(json.loads(updated.master_json))}" + } + + # New incident — create record + incident = IncidentMasterData( + incident_id=incident_id, + master_json=json.dumps(extracted), + transcript_text=input_text, + location_lat=location_lat, + location_lng=location_lng, + ) + saved = create_incident(db, incident) + print(f"[DATA LAKE] Stored incident {incident_id} with {len(extracted)} fields") + + return { + "incident_id": incident_id, + "status": "created", + "fields_extracted": len(extracted), + "master_json": extracted, + "message": f"Incident data stored. Generate PDFs with POST /incidents/{incident_id}/generate/{{template_id}}" + } + + +# ── Generate PDF from stored data ──────────────────────── + +@router.post("/{incident_id}/generate/{template_id}") +def generate_pdf_from_lake( + incident_id: str, + template_id: int, + db: Session = Depends(get_db) +): + """ + Generates a PDF for any agency template from the stored Master Incident Data Lake. + Supports dynamic multi-template generation from a single incident record — + Record Once, Report Everywhere. + """ + incident = get_incident(db, incident_id) + if not incident: + raise AppError(f"Incident {incident_id} not found in data lake", status_code=404) + + template = get_template(db, template_id) + if not template: + raise AppError(f"Template {template_id} not found", status_code=404) + + if not os.path.exists(template.pdf_path): + raise AppError(f"Template PDF not found on disk: {template.pdf_path}", status_code=404) + + print(f"[DATA LAKE] Generating '{template.name}' from incident {incident_id}") + + master_data = json.loads(incident.master_json) + tpl_fields = list(template.fields.keys()) if isinstance(template.fields, dict) else template.fields + + # Map stored Data Lake fields to this template's fields + mapped_data = {k: master_data.get(k) for k in tpl_fields if master_data.get(k) is not None} + + print(f"[DATA LAKE] Template needs {len(tpl_fields)} fields, matched {len(mapped_data)}") + + # Fill PDF + filler = Filler() + try: + output_path = filler.fill_form_with_data( + pdf_form=template.pdf_path, + data=mapped_data + ) + except Exception as e: + raise AppError(f"PDF generation failed: {str(e)}", status_code=500) + + if not output_path or not os.path.exists(output_path): + raise AppError("PDF generation produced no output", status_code=500) + + # Save submission record + submission = FormSubmission( + template_id=template_id, + input_text=f"[DATA LAKE] {incident_id}", + output_pdf_path=output_path + ) + saved = create_form(db, submission) + + return { + "incident_id": incident_id, + "template_id": template_id, + "template_name": template.name, + "submission_id": saved.id, + "download_url": f"/forms/download/{saved.id}", + "fields_matched": len(mapped_data), + "fields_total": len(tpl_fields), + "message": "PDF generated from Master Data Lake." + } + + +# ── Get incident data ──────────────────────────────────── + +@router.get("/{incident_id}") +def get_incident_data(incident_id: str, db: Session = Depends(get_db)): + """Get stored master JSON for an incident.""" + incident = get_incident(db, incident_id) + if not incident: + raise AppError(f"Incident {incident_id} not found", status_code=404) + return { + "incident_id": incident.incident_id, + "master_json": json.loads(incident.master_json), + "transcript": incident.transcript_text, + "location": { + "lat": incident.location_lat, + "lng": incident.location_lng + } if incident.location_lat else None, + "created_at": incident.created_at, + "updated_at": incident.updated_at + } + + +# ── List all incidents ─────────────────────────────────── + +@router.get("") +def list_incidents(db: Session = Depends(get_db)): + """List all incidents in the data lake.""" + incidents = get_all_incidents(db) + return [ + { + "incident_id": i.incident_id, + "fields_count": len(json.loads(i.master_json)), + "created_at": i.created_at, + "location": {"lat": i.location_lat, "lng": i.location_lng} if i.location_lat else None + } + for i in incidents + ] + + +# ── Narrative generation ───────────────────────────────── + +@router.post("/{incident_id}/narrative") +def generate_narrative(incident_id: str, db: Session = Depends(get_db)): + """ + Generate a legally coherent narrative report from stored incident data. + For insurance claims, court documents, after-action reports. + Uses the LLM to write prose — not fill fields. + """ + incident = get_incident(db, incident_id) + if not incident: + raise AppError(f"Incident {incident_id} not found", status_code=404) + + master_data = json.loads(incident.master_json) + fields_summary = "\n".join([f"- {k}: {v}" for k, v in master_data.items() if v]) + + narrative_prompt = f"""You are a professional incident report writer for emergency services. +Based on the following structured incident data, write a clear, factual, legally coherent +narrative report suitable for insurance claims and court documentation. + +Incident ID: {incident_id} +Date/Time: {incident.created_at} +Original Transcript: {incident.transcript_text} + +Extracted Data: +{fields_summary} + +Write a professional narrative report in 3-4 paragraphs covering: +1. Incident summary (what happened, when, where) +2. Response and actions taken +3. Outcome and follow-up required + +Use formal language appropriate for legal documentation.""" + + try: + import requests + response = requests.post( + "http://localhost:11434/api/generate", + json={ + "model": "mistral", + "prompt": narrative_prompt, + "stream": False + }, + timeout=120 + ) + narrative = response.json().get("response", "").strip() + except Exception as e: + raise AppError(f"Narrative generation failed: {str(e)}", status_code=500) + + return { + "incident_id": incident_id, + "narrative": narrative, + "format": "markdown", + "generated_at": datetime.utcnow().isoformat() + } \ No newline at end of file diff --git a/api/routes/templates.py b/api/routes/templates.py index 5c2281b..9419ae6 100644 --- a/api/routes/templates.py +++ b/api/routes/templates.py @@ -1,16 +1,89 @@ -from fastapi import APIRouter, Depends +import os +import shutil +import uuid +from fastapi import APIRouter, Depends, UploadFile, File, Form from sqlmodel import Session from api.deps import get_db -from api.schemas.templates import TemplateCreate, TemplateResponse -from api.db.repositories import create_template +from api.schemas.templates import TemplateResponse +from api.db.repositories import create_template, get_all_templates from api.db.models import Template -from src.controller import Controller +from api.errors.base import AppError router = APIRouter(prefix="/templates", tags=["templates"]) +# Save directly into src/inputs/ — stable location, won't get wiped +TEMPLATES_DIR = os.path.join("src", "inputs") +os.makedirs(TEMPLATES_DIR, exist_ok=True) + + @router.post("/create", response_model=TemplateResponse) -def create(template: TemplateCreate, db: Session = Depends(get_db)): - controller = Controller() - template_path = controller.create_template(template.pdf_path) - tpl = Template(**template.model_dump(exclude={"pdf_path"}), pdf_path=template_path) - return create_template(db, tpl) \ No newline at end of file +async def create( + name: str = Form(...), + file: UploadFile = File(...), + db: Session = Depends(get_db) +): + # Validate PDF + if not file.filename.endswith(".pdf"): + raise AppError("Only PDF files are allowed", status_code=400) + + # Save uploaded file with unique name into src/inputs/ + unique_name = f"{uuid.uuid4().hex}_{file.filename}" + save_path = os.path.join(TEMPLATES_DIR, unique_name) + + with open(save_path, "wb") as f: + shutil.copyfileobj(file.file, f) + + # Extract fields using commonforms + pypdf + # Store as simple list of field name strings — what Filler expects + try: + from commonforms import prepare_form + from pypdf import PdfReader + + # Read real field names directly from original PDF + # Use /T (internal name) as both key and label + # Real names like "JobTitle", "Phone Number" are already human-readable + reader = PdfReader(save_path) + raw_fields = reader.get_fields() or {} + + fields = {} + for internal_name, field_data in raw_fields.items(): + # Use /TU tooltip if available, otherwise prettify /T name + label = None + if isinstance(field_data, dict): + label = field_data.get("/TU") + if not label: + # Prettify: "JobTitle" → "Job Title", "DATE7_af_date" → "Date" + import re + label = re.sub(r'([a-z])([A-Z])', r'\1 \2', internal_name) + label = re.sub(r'_af_.*$', '', label) # strip "_af_date" suffix + label = label.replace('_', ' ').strip().title() + fields[internal_name] = label + + except Exception as e: + print(f"Field extraction failed: {e}") + fields = [] + + # Save to DB + tpl = Template(name=name, pdf_path=save_path, fields=fields) + return create_template(db, tpl) + + +@router.get("", response_model=list[TemplateResponse]) +def list_templates( + limit: int = 100, + offset: int = 0, + db: Session = Depends(get_db) +): + return get_all_templates(db, limit=limit, offset=offset) + + +@router.get("/{template_id}", response_model=TemplateResponse) +def get_template_by_id( + template_id: int, + db: Session = Depends(get_db) +): + from api.db.repositories import get_template + tpl = get_template(db, template_id) + if not tpl: + raise AppError("Template not found", status_code=404) + return tpl \ No newline at end of file diff --git a/api/routes/transcribe.py b/api/routes/transcribe.py new file mode 100644 index 0000000..41783fa --- /dev/null +++ b/api/routes/transcribe.py @@ -0,0 +1,77 @@ + +from fastapi import APIRouter, UploadFile, File, Query +from fastapi.responses import JSONResponse +from api.errors.base import AppError +from src.transcriber import transcribe_audio + +router = APIRouter(prefix="/transcribe", tags=["transcription"]) + +ALLOWED_EXTENSIONS = {".mp3", ".mp4", ".wav", ".m4a", ".ogg", ".webm", ".flac"} +MAX_FILE_SIZE = 50 * 1024 * 1024 # 50MB + + +@router.post("") +async def transcribe( + file: UploadFile = File(...), + language: str = Query( + default=None, + description="Optional language code e.g. 'en', 'fr', 'es'. " + "Leave empty for auto-detection." + ) +): + """ + Transcribe an audio file to text using faster-whisper. + + Upload any audio file (wav, mp3, m4a, webm, ogg). + Returns transcript text ready to pass directly into POST /forms/fill. + + Works CPU-only — no GPU required. + Typical transcription time: 2-5s for a 1-minute recording. + + Example workflow: + 1. POST /transcribe → get transcript + 2. POST /forms/fill → fill PDF from transcript + """ + # Validate file extension + from pathlib import Path + ext = Path(file.filename or "").suffix.lower() + if ext not in ALLOWED_EXTENSIONS: + raise AppError( + f"Unsupported file type '{ext}'. " + f"Supported: {', '.join(sorted(ALLOWED_EXTENSIONS))}", + status_code=422 + ) + + # Read and validate file size + file_bytes = await file.read() + if len(file_bytes) > MAX_FILE_SIZE: + raise AppError( + f"File too large ({len(file_bytes) // (1024*1024)}MB). " + "Maximum allowed size is 50MB.", + status_code=413 + ) + + if len(file_bytes) == 0: + raise AppError("Uploaded file is empty.", status_code=422) + + try: + result = transcribe_audio( + file_bytes=file_bytes, + filename=file.filename or "audio.wav", + language=language or None + ) + except RuntimeError as e: + raise AppError(str(e), status_code=503) + except Exception as e: + raise AppError( + f"Transcription failed: {str(e)}", + status_code=500 + ) + + return { + "transcript": result["transcript"], + "language": result["language"], + "language_probability": result["language_probability"], + "duration_seconds": result["duration"], + "hint": "Pass 'transcript' directly as 'input_text' to POST /forms/fill" + } \ No newline at end of file diff --git a/api/schemas/forms.py b/api/schemas/forms.py index 3cce650..ae8c38b 100644 --- a/api/schemas/forms.py +++ b/api/schemas/forms.py @@ -1,15 +1,65 @@ from pydantic import BaseModel +from typing import Optional +from datetime import datetime + class FormFill(BaseModel): template_id: int input_text: str + class Config: + from_attributes = True + class FormFillResponse(BaseModel): id: int template_id: int input_text: str output_pdf_path: str + created_at: datetime + + class Config: + from_attributes = True + + +# ── Batch schemas — closes #156 ─────────────────────────────── + +class BatchFormFill(BaseModel): + """ + Request body for POST /forms/fill/batch. + One transcript + multiple template IDs → fills all PDFs in one request. + """ + input_text: str + template_ids: list[int] + + class Config: + from_attributes = True + + +class BatchResultItem(BaseModel): + """ + Per-template result in a batch fill response. + """ + template_id: int + template_name: str + success: bool + submission_id: Optional[int] = None + download_url: Optional[str] = None + error: Optional[str] = None + + class Config: + from_attributes = True + + +class BatchFormFillResponse(BaseModel): + """ + Response body for POST /forms/fill/batch. + Partial failures preserved — one failure never aborts the batch. + """ + total: int + succeeded: int + failed: int + results: list[BatchResultItem] class Config: from_attributes = True \ No newline at end of file diff --git a/docs/SETUP.md b/docs/SETUP.md new file mode 100644 index 0000000..3d406de --- /dev/null +++ b/docs/SETUP.md @@ -0,0 +1,464 @@ +# 🔥 FireForm — Setup & Usage Guide + +This guide covers how to install, run, and use FireForm locally on Windows, Linux, and macOS. + +--- + +## 📋 Prerequisites + +| Tool | Version | Purpose | +|------|---------|---------| +| Python | 3.11+ | Backend runtime | +| Ollama | 0.17.7+ | Local LLM server | +| Mistral 7B | latest | AI extraction model | +| Git | any | Clone the repository | + +--- + +## 🪟 Windows + +### 1. Clone the repository +```cmd +git clone https://github.com/fireform-core/FireForm.git +cd FireForm +``` + +### 2. Create and activate virtual environment +```cmd +python -m venv venv +venv\Scripts\activate +``` + +### 3. Install dependencies +```cmd +pip install -r requirements.txt +``` + +### 4. Install and start Ollama +Download Ollama from https://ollama.com/download/windows + +Then pull the Mistral model: +```cmd +ollama pull mistral +ollama serve +``` + +> Ollama runs on `http://localhost:11434` by default. Keep this terminal open. + +### 5. Initialize the database +```cmd +python -m api.db.init_db +``` + +### 6. Start the API server +```cmd +uvicorn api.main:app --reload +``` + +API is now running at `http://127.0.0.1:8000` + +### 7. Start the frontend +Open a new terminal: +```cmd +cd frontend +python -m http.server 3000 +``` + +Open `http://localhost:3000` in your browser. + +--- + +## 🐧 Linux (Ubuntu/Debian) + +### 1. Clone and enter the repository +```bash +git clone https://github.com/fireform-core/FireForm.git +cd FireForm +``` + +### 2. Create and activate virtual environment +```bash +python3 -m venv venv +source venv/bin/activate +``` + +### 3. Install dependencies +```bash +pip install -r requirements.txt +``` + +### 4. Install and start Ollama +```bash +curl -fsSL https://ollama.com/install.sh | sh +ollama pull mistral +ollama serve & +``` + +### 5. Initialize the database +```bash +python -m api.db.init_db +``` + +### 6. Start the API server +```bash +uvicorn api.main:app --reload +``` + +### 7. Start the frontend +```bash +cd frontend +python3 -m http.server 3000 +``` + +--- + +## 🍎 macOS + +### 1. Clone and enter the repository +```bash +git clone https://github.com/fireform-core/FireForm.git +cd FireForm +``` + +### 2. Create and activate virtual environment +```bash +python3 -m venv venv +source venv/bin/activate +``` + +### 3. Install dependencies +```bash +pip install -r requirements.txt +``` + +### 4. Install and start Ollama +Download from https://ollama.com/download/mac or: +```bash +brew install ollama +ollama pull mistral +ollama serve & +``` + +### 5. Initialize the database +```bash +python -m api.db.init_db +``` + +### 6. Start the API server +```bash +uvicorn api.main:app --reload +``` + +### 7. Start the frontend +```bash +cd frontend +python3 -m http.server 3000 +``` + +--- + +## 🖥️ Using the Frontend + +Once everything is running, open `http://localhost:3000` in your browser. + +### Step 1 — Upload a PDF template +- Click **"Choose File"** and select any fillable PDF form +- Enter a name for the template +- Click **"Upload Template"** + +FireForm will automatically extract all form field names and their human-readable labels. + +### Step 2 — Fill the form +- Select your uploaded template from the dropdown +- In the text box, describe the incident or enter the information in natural language: + +``` +Employee name is John Smith. Employee ID is EMP-2024-789. +Job title is Firefighter Paramedic. Location is Station 12 Sacramento. +Department is Emergency Medical Services. Supervisor is Captain Rodriguez. +Phone number is 916-555-0147. +``` + +- Click **"Fill Form"** + +FireForm sends one request to Ollama (Mistral) which extracts all fields at once and returns structured JSON. + + +### Batch fill — multiple agency forms at once + +Switch to **BATCH** mode in the sidebar to fill multiple templates simultaneously from one transcript: + +1. Click **BATCH** toggle in the sidebar +2. Check all agency templates you want to fill +3. Enter one incident description +4. Click **⚡ FILL N FORMS** + +FireForm runs a single LLM call for the entire batch and returns individual download links for each filled PDF. One failed template never aborts the rest. + +--- +### Step 3 — Download the filled PDF +- Click **"Download PDF"** to save the completed form + +--- + +## ✅ Supported PDF Field Types + +FireForm supports all common fillable PDF field types: + +| Field Type | Description | Example | +|------------|-------------|---------| +| Text | Plain text input | Name, ID, Notes | +| Checkbox | Boolean tick box | Married ✓ | +| Radio button | Single selection from options | Gender: Male / Female | +| Dropdown | Single select list | City | +| Multi-select | Multiple select list | Language | + +**Checkbox and radio button filling:** +FireForm automatically detects the field type from the PDF annotation flags (`FT` and `Ff`) and writes the correct PDF value format. PDF checkboxes require named values like `/Yes` or `/Off` — not plain strings. FireForm reads the PDF's own appearance stream (`AP.N`) to find the exact on-state name used by each form, so it works correctly with any PDF regardless of internal naming conventions. + +LLM outputs like `"yes"`, `"true"`, `"x"`, `"1"`, `"checked"` all resolve to the correct checked state. Outputs like `"no"`, `"false"`, `"0"`, `""` resolve to unchecked. + +--- + +## 🤖 How AI Extraction Works + +FireForm uses a **batch extraction** approach: + +``` +Traditional approach (slow): FireForm approach (fast): + Field 1 → Ollama call All fields → 1 Ollama call + Field 2 → Ollama call Mistral returns JSON with all values + Field 3 → Ollama call Parse → fill PDF + ...N calls total 1 call total (O(1)) +``` + +Field names are automatically read from the PDF's annotations and converted to human-readable labels before being sent to Mistral — so the model understands what each field means regardless of internal PDF naming conventions like `textbox_0_0`. + +**Example extraction:** +```json +{ + "NAME/SID": "John Smith", + "JobTitle": "Firefighter Paramedic", + "Department": "Emergency Medical Services", + "Phone Number": "916-555-0147", + "email": null +} +``` + +--- + +## 🧪 Running Tests + +```bash +python -m pytest tests/ -v +``` + +Expected output: **70 passed** + +See [TESTING.md](TESTING.md) for full test coverage details. + +--- + +## 🔧 Environment Variables + +| Variable | Default | Description | +|----------|---------|-------------| +| `OLLAMA_HOST` | `http://localhost:11434` | Ollama server URL | + +To use a remote Ollama instance: +```bash +export OLLAMA_HOST=http://your-server:11434 # Linux/Mac +set OLLAMA_HOST=http://your-server:11434 # Windows +``` + +--- + +## 🐳 Docker (Coming Soon) + +Docker support is in progress. See [docker.md](docker.md) for current status. + +--- + +## ❓ Troubleshooting + +**`Form data requires python-multipart`** +```bash +pip install python-multipart +``` + +**`ModuleNotFoundError: No module named 'pypdf'`** +```bash +pip install pypdf +``` + +**`Could not connect to Ollama`** +- Make sure `ollama serve` is running +- Check Ollama is on port 11434: `curl http://localhost:11434` + +**`NameError: name 'Union' is not defined`** +- Pull latest changes: `git pull origin main` +- This bug is fixed in the current version + +**Tests fail with `ModuleNotFoundError: No module named 'api'`** +- Use `python -m pytest` instead of `pytest` + +--- + +## 🗄️ Master Incident Data Lake + +FireForm now ships with a persistent **Master Incident Data Lake** — a foundational backend architecture that decouples voice extraction from rigid single-PDF workflows, enabling the *"Record Once. Report Everywhere."* paradigm. + +### What is the Data Lake? + +Instead of extracting from a transcript → filling one PDF → discarding all data, FireForm now: + +1. Extracts **all spoken intelligence** into a permanent, schema-less JSON record linked to a unique **Incident ID** (`INC-YYYY-MMDD-HHMM`). +2. Stores it in the database — independently of any PDF template. +3. Lets any officer, at any time, generate a filled PDF for **any registered agency template** from that same stored record — with zero new LLM calls. + +``` +Old approach: + Transcript → LLM → PDF → ❌ Data discarded + +Master Data Lake approach: + Transcript → LLM → Master JSON (persisted) → PDF A + → PDF B + → PDF C (any template, any time) +``` + +--- + +### Data Lake Workflow + +#### Step 1 — Record an Incident + +Enter your incident description in the text box and click **"Save to Data Lake"** (or use the API directly): + +``` +POST /incidents/extract?input_text=&incident_id= +``` + +If no `incident_id` is provided, one is auto-generated. A unique Incident ID is returned: + +```json +{ + "incident_id": "INC-2026-0401-0912", + "status": "created", + "fields_extracted": 7 +} +``` + +> **Tip:** Copy and save your Incident ID. You will need it to append data or generate PDFs. + +--- + +#### Step 2 — Append Data (Collaborative Reporting) + +Multiple officers can contribute to the same incident record by passing the same `incident_id`: + +``` +POST /incidents/extract?input_text=&incident_id=INC-2026-0401-0912 +``` + +FireForm's **Collaborative Consensus Merge** engine handles conflicts intelligently: + +| Scenario | Behaviour | +|----------|-----------| +| New officer sends `null` for a field that already has data | Existing value is **protected** (not overwritten) | +| New officer adds a field not previously seen | Field is **added** to the Data Lake | +| Both officers mention `Notes` or `Description` | Values are **appended** with a timestamped `[UPDATE]` tag | +| New officer corrects a non-null field with a new value | Value is **updated** | + +The response will include `"status": "merged"`. + +--- + +#### Step 3 — Generate a PDF for Any Agency Template + +Once the incident is stored, generate a filled PDF for any uploaded template: + +``` +POST /incidents/{incident_id}/generate/{template_id} +``` + +Example: +``` +POST /incidents/INC-2026-0401-0912/generate/3 +``` + +FireForm maps the stored Data Lake JSON to the selected template's fields and returns a download link: + +```json +{ + "incident_id": "INC-2026-0401-0912", + "template_name": "Fire Department Report", + "submission_id": 12, + "download_url": "/forms/download/12", + "fields_matched": 6, + "fields_total": 8 +} +``` + +You can call this endpoint multiple times with different `template_id` values — one incident record, unlimited reports. + +--- + +#### Step 4 — Inspect the Data Lake + +Retrieve the full raw master JSON at any time: + +``` +GET /incidents/{incident_id} +``` + +List all stored incidents: + +``` +GET /incidents +``` + +--- + +### Data Lake API Reference + +| Method | Endpoint | Description | +|--------|----------|-------------| +| `POST` | `/incidents/extract` | Extract transcript → store in Data Lake | +| `GET` | `/incidents` | List all stored incidents | +| `GET` | `/incidents/{id}` | Retrieve full master JSON for one incident | +| `POST` | `/incidents/{id}/generate/{template_id}` | Generate a PDF from stored data | + +--- + +### Environment Variables (Updated) + +| Variable | Default | Description | +|----------|---------|-------------| +| `OLLAMA_HOST` | `http://localhost:11434` | Ollama server URL | +| `OLLAMA_TIMEOUT` | `300` | LLM request timeout in seconds (increase for slow hardware) | + +To customise: +```bash +export OLLAMA_HOST=http://your-server:11434 # Linux/Mac +export OLLAMA_TIMEOUT=300 # Linux/Mac + +set OLLAMA_HOST=http://your-server:11434 # Windows +set OLLAMA_TIMEOUT=300 # Windows +``` + +--- + +### Running Data Lake Tests + +The Data Lake test suite uses an in-memory SQLite database and mocks the LLM — **no Ollama instance required**: + +```bash +python -m pytest tests/test_incidents.py -v +``` + +Expected output: **13 passed** + +Full test suite: +```bash +python -m pytest tests/ -v +``` \ No newline at end of file diff --git a/frontend/index.html b/frontend/index.html new file mode 100644 index 0000000..411c120 --- /dev/null +++ b/frontend/index.html @@ -0,0 +1,1615 @@ + + + + + + + FireForm — Report Once, File Everywhere + + + + + +
+
+ + + +
+ + +
+
+
UN Digital Public Good · GSoC 2026
+

REPORT
ONCE.

+

Describe any incident in plain language. FireForm uses a locally-running AI to extract every + relevant detail and auto-fill all required agency forms — instantly and privately.

+
+ +
+
+
1
+
+
Upload Template
+
Any fillable PDF form
+
+
+
+
2
+
+
Select Template(s)
+
Single or multi-agency batch
+
+
+
+
3
+
+
Describe Incident
+
Plain language report
+
+
+
+
4
+
+
Download PDF
+
All fields auto-filled
+
+
+
+ +
+
← Select a template from the sidebar
+
+ Incident Description * + 0 chars +
+
+ +
+
Click mic to record incident report
+
+ Transcribing audio... +
+
+
+ + +
+ +
Runs via Ollama locally.
No data leaves your machine.
+
+
+
+ +
Mistral is extracting data and filling your form...
+
+
+
+
✓ FORM FILLED SUCCESSFULLY
+ +
+
+
+
+
+
+
+ +
+
+
Session History
+
0 submissions
+
+
+
No submissions yet this session.
+
+
+
+
+ + + + + \ No newline at end of file diff --git a/requirements.txt b/requirements.txt index eaa6c81..405c441 100644 --- a/requirements.txt +++ b/requirements.txt @@ -9,4 +9,5 @@ sqlmodel pytest httpx numpy<2 -ollama \ No newline at end of file +ollama +faster-whisper \ No newline at end of file diff --git a/src/filler.py b/src/filler.py index e31e535..3e2ab66 100644 --- a/src/filler.py +++ b/src/filler.py @@ -1,16 +1,144 @@ -from pdfrw import PdfReader, PdfWriter +from pdfrw import PdfReader, PdfWriter, PdfName from src.llm import LLM from datetime import datetime +TRUTHY_VALUES = {"yes", "true", "1", "on", "checked", "x", "selected", "male", "female"} +FALSY_VALUES = {"no", "false", "0", "off", "unchecked", "", "none", "null"} + + +def _resolve_checkbox_value(raw_value, annot): + """Convert LLM string → correct PDF checkbox value (/Yes or /Off).""" + normalized = str(raw_value).strip().lower() + is_checked = normalized in TRUTHY_VALUES + if is_checked: + try: + if annot.AP and annot.AP.N: + for key in annot.AP.N.keys(): + clean = str(key).strip("/") + if clean.lower() not in ("off", "false", "0", "length", + "subtype", "bbox", "resources"): + return PdfName(clean) + except Exception: + pass + return PdfName("Yes") + return PdfName("Off") + + +def _resolve_radio_kid(raw_value, kid_index, annot): + """Determine if THIS radio kid should be selected.""" + normalized = str(raw_value).strip().lower() + try: + parent = annot.Parent + if parent and parent.Opt: + opts = [str(o).strip("()").lower() for o in parent.Opt] + if kid_index < len(opts) and opts[kid_index] == normalized: + if annot.AP and annot.AP.N: + for key in annot.AP.N.keys(): + clean = str(key).strip("/") + if clean.lower() not in ("off", "false", "0"): + return PdfName(clean) + return PdfName(str(kid_index)) + except Exception: + pass + return PdfName("Off") + + +def _get_field_type(annot): + """Return 'text', 'checkbox', 'radio', 'dropdown', 'pushbutton', or 'other'.""" + ft = str(annot.FT).strip("/") if annot.FT else "" + if ft == "Btn": + try: + ff = int(str(annot.Ff)) if annot.Ff else 0 + if ff & (1 << 16): + return "pushbutton" + if ff & (1 << 15): + return "radio" + except Exception: + pass + return "checkbox" + elif ft == "Tx": + return "text" + elif ft == "Ch": + return "dropdown" + return "other" + + +def _fill_annotation(annot, raw_value) -> str: + """Write correct value to annotation based on field type and return the written value for logging.""" + field_type = _get_field_type(annot) + written_val = "" + + if field_type == "checkbox": + annot.V = _resolve_checkbox_value(raw_value, annot) + annot.AS = annot.V + written_val = str(annot.V) + + elif field_type == "radio": + if annot.Kids: + normalized = str(raw_value).strip().lower() + selected_index = None + try: + if annot.Opt: + opts = [str(o).strip("()").lower() for o in annot.Opt] + if normalized in opts: + selected_index = opts.index(normalized) + except Exception: + pass + + for i, kid in enumerate(annot.Kids): + kid_on_key = None + try: + if kid.AP and kid.AP.N: + for key in kid.AP.N.keys(): + clean = str(key).strip("/") + if clean.lower() not in ("off", "false", "0"): + kid_on_key = clean + break + except Exception: + pass + + # Match by explicit /Opt index, OR by direct match to the internal graphic key! + if (selected_index is not None and i == selected_index) or \ + (kid_on_key and normalized in kid_on_key.lower()): + on_val = PdfName(kid_on_key if kid_on_key else str(i)) + kid.AS = on_val + annot.V = on_val + written_val = str(on_val) + else: + kid.AS = PdfName("Off") + else: + annot.V = _resolve_checkbox_value(raw_value, annot) + annot.AS = annot.V + written_val = str(annot.V) + + elif field_type == "pushbutton": + written_val = "Skipped" + + elif field_type == "dropdown": + annot.V = "" if raw_value is None else str(raw_value) + written_val = str(annot.V) + + else: + # Plain text — never write literal "None" + annot.V = "" if raw_value is None else str(raw_value) + annot.AP = None # Moved inside text block! Checkboxes preserve appearance! + written_val = str(annot.V) + + return written_val + + class Filler: def __init__(self): pass def fill_form(self, pdf_form: str, llm: LLM): """ - Fill a PDF form with values from user_input using LLM. - Fields are filled in the visual order (top-to-bottom, left-to-right). + Fill a PDF form using LLM extraction. + Uses KEY-BASED matching — field name from PDF matched to + extracted JSON key. This ensures correct data goes to + correct field regardless of PDF field order. + Falls back to positional if key not found in extraction. """ output_pdf = ( pdf_form[:-4] @@ -19,34 +147,131 @@ def fill_form(self, pdf_form: str, llm: LLM): + "_filled.pdf" ) - # Generate dictionary of answers from your original function t2j = llm.main_loop() - textbox_answers = t2j.get_data() # This is a dictionary + extracted = t2j.get_data() # dict: {field_name: value} - answers_list = list(textbox_answers.values()) + print(f"[FILLER] Extracted {len(extracted)} fields:") + for k, v in extracted.items(): + print(f" {k}: {v}") - # Read PDF pdf = PdfReader(pdf_form) - # Loop through pages + processed_parents = set() + for page in pdf.pages: if page.Annots: - sorted_annots = sorted( - page.Annots, key=lambda a: (-float(a.Rect[1]), float(a.Rect[0])) - ) - - i = 0 - for annot in sorted_annots: - if annot.Subtype == "/Widget" and annot.T: - if i < len(answers_list): - annot.V = f"{answers_list[i]}" - annot.AP = None - i += 1 + for annot in page.Annots: + if annot.Subtype != "/Widget": + continue + + # Direct field (has its own T key) + if annot.T: + # Clean field key — strip pdfrw parentheses + field_key = annot.T.strip("()") + + # Try exact key match first + raw = extracted.get(field_key) + + # Try case-insensitive match if exact fails + if raw is None: + for k, v in extracted.items(): + if k.lower() == field_key.lower(): + raw = v + break + + if raw is not None: + written_val = _fill_annotation(annot, raw) + print(f" [FILLER] Filling '{field_key}' = {raw} → {written_val} \u2713") else: - # Stop if we run out of answers - break + print(f" [FILLER] No match for '{field_key}' — leaving empty") - PdfWriter().write(output_pdf, pdf) + # Radio button kid (T key is on the parent) + elif annot.Parent and annot.Parent.T: + parent = annot.Parent + if id(parent) in processed_parents: + continue + processed_parents.add(id(parent)) + + field_key = parent.T.strip("()") + raw = extracted.get(field_key) + if raw is None: + for k, v in extracted.items(): + if k.lower() == field_key.lower(): + raw = v + break + + if raw is not None: + written_val = _fill_annotation(parent, raw) + print(f" [FILLER] Filling '{field_key}' = {raw} → {written_val} \u2713") + else: + print(f" [FILLER] No match for parent '{field_key}' — leaving empty") - # Your main.py expects this function to return the path + PdfWriter().write(output_pdf, pdf) + print("\nlog extracted successfully") + print(f"along with what it extracted accordingly, pdf file : {output_pdf}") return output_pdf + + def fill_form_with_data(self, pdf_form: str, data: dict) -> str: + """ + Fill a PDF form with pre-extracted data dictionary. + Used by batch endpoint — NO LLM call. + Key-based matching with case-insensitive fallback. + """ + print(f"[log extracted successfully] Found {len(data)} fields mapped from Data Lake.") + + output_pdf = ( + pdf_form[:-4] + + "_" + + datetime.now().strftime("%Y%m%d_%H%M%S") + + "_filled.pdf" + ) + + pdf = PdfReader(pdf_form) + + processed_parents = set() + + for page in pdf.pages: + if page.Annots: + for annot in page.Annots: + if annot.Subtype != "/Widget": + continue + + if annot.T: + field_key = annot.T.strip("()") + + # Exact match + raw = data.get(field_key) + + # Case-insensitive fallback + if raw is None: + for k, v in data.items(): + if k.lower() == field_key.lower(): + raw = v + break + + if raw is not None: + written_val = _fill_annotation(annot, raw) + print(f" [FILLER] Filling '{field_key}' = {raw} → {written_val} \u2713") + + elif annot.Parent and annot.Parent.T: + parent = annot.Parent + if id(parent) in processed_parents: + continue + processed_parents.add(id(parent)) + + field_key = parent.T.strip("()") + raw = data.get(field_key) + if raw is None: + for k, v in data.items(): + if k.lower() == field_key.lower(): + raw = v + break + + if raw is not None: + written_val = _fill_annotation(parent, raw) + print(f" [FILLER] Filling '{field_key}' = {raw} → {written_val} \u2713") + + PdfWriter().write(output_pdf, pdf) + print("\nlog extracted successfully") + print(f"along with what it extracted accordingly, pdf file : {output_pdf}") + return output_pdf \ No newline at end of file diff --git a/src/llm.py b/src/llm.py index 70937f9..ad331e3 100644 --- a/src/llm.py +++ b/src/llm.py @@ -1,14 +1,19 @@ import json import os +import time import requests class LLM: def __init__(self, transcript_text=None, target_fields=None, json=None): + """ + target_fields: dict or list containing the template field names to extract + (dict format: {"field_name": "human_label"}, list format: ["field_name1", "field_name2"]) + """ if json is None: json = {} self._transcript_text = transcript_text # str - self._target_fields = target_fields # List, contains the template field. + self._target_fields = target_fields # dict or list self._json = json # dictionary def type_check_all(self): @@ -17,76 +22,236 @@ def type_check_all(self): f"ERROR in LLM() attributes ->\ Transcript must be text. Input:\n\ttranscript_text: {self._transcript_text}" ) - elif type(self._target_fields) is not list: + if not isinstance(self._target_fields, (list, dict)): raise TypeError( f"ERROR in LLM() attributes ->\ - Target fields must be a list. Input:\n\ttarget_fields: {self._target_fields}" + Target fields must be a list or dict. Input:\n\ttarget_fields: {self._target_fields}" ) - def build_prompt(self, current_field): + def build_batch_prompt(self) -> str: """ - This method is in charge of the prompt engineering. It creates a specific prompt for each target field. - @params: current_field -> represents the current element of the json that is being prompted. + Build a single prompt that extracts fields at once. + Supports BOTH template-guided and pure schema-less dynamic extraction! """ - prompt = f""" - SYSTEM PROMPT: - You are an AI assistant designed to help fillout json files with information extracted from transcribed voice recordings. - You will receive the transcription, and the name of the JSON field whose value you have to identify in the context. Return - only a single string containing the identified value for the JSON field. - If the field name is plural, and you identify more than one possible value in the text, return both separated by a ";". - If you don't identify the value in the provided text, return "-1". - --- - DATA: - Target JSON field to find in text: {current_field} - - TEXT: {self._transcript_text} - """ + if not self._target_fields: + # PURE SCHEMA-LESS: No templates exist, purely ad-hoc extraction! + prompt = f"""You are an advanced data extraction engine. +Extract every meaningful piece of information from the transcript below. + +RULES: +1. Return ONLY a valid JSON object — no explanation, no markdown, no extra text +2. You MUST dynamically invent descriptive JSON keys for every critical detail (e.g. "Injuries", "Weapons", "SuspectName", "Location"). +3. Always pair the invented key with its exact value from the transcript. +4. For multiple values, use a semicolon-separated string: "Name1; Name2" + +TRANSCRIPT: +{self._transcript_text} + +JSON:""" + return prompt + + # TEMPLATE-GUIDED + DYNAMIC EXTRACTION + if isinstance(self._target_fields, dict): + fields_lines = "\n".join( + f' "{k}": null // {v if v and v != k else k}' + for k, v in self._target_fields.items() + ) + else: + fields_lines = "\n".join( + f' "{f}": null' + for f in self._target_fields + ) + + prompt = f"""You are filling out an official form. Extract values from the transcript below. + +FORM FIELDS (each line: "internal_key": null // visible label on form): +{{ +{fields_lines} +}} + +RULES: +1. Return ONLY a valid JSON object — no explanation, no markdown, no extra text +2. Use the visible label (after //) to understand what each field means +3. Fill each key with the matching value from the transcript +4. If a value is not found in the transcript, use null +5. Never invent or guess values not present in the transcript +6. For multiple values (e.g. multiple victims), use a semicolon-separated string: "Name1; Name2" +7. Distinguish roles carefully: Officer/Employee is NOT the same as Victim or Suspect +8. IMPORTANT: You MUST recursively extract any other critical details found in the transcript by inventing your own descriptive JSON keys (e.g. "Weapon": "Glock", "Injury": "Broken Leg"). + +TRANSCRIPT: +{self._transcript_text} + +JSON:""" return prompt - def main_loop(self): - # self.type_check_all() - for field in self._target_fields.keys(): - prompt = self.build_prompt(field) - # print(prompt) - # ollama_url = "http://localhost:11434/api/generate" - ollama_host = os.getenv("OLLAMA_HOST", "http://localhost:11434").rstrip("/") - ollama_url = f"{ollama_host}/api/generate" - - payload = { - "model": "mistral", - "prompt": prompt, - "stream": False, # don't really know why --> look into this later. - } + def build_prompt(self, current_field: str) -> str: + """ + Legacy single-field prompt — kept for backward compatibility. + Used as fallback if batch parsing fails. + """ + field_lower = current_field.lower() + is_plural = current_field.endswith('s') and not current_field.lower().endswith('ss') - try: - response = requests.post(ollama_url, json=payload) + if any(w in field_lower for w in ['officer', 'employee', 'dispatcher', 'caller', 'reporting', 'supervisor']): + role_guidance = """ +ROLE: Extract the PRIMARY OFFICER/EMPLOYEE/DISPATCHER +- This is typically the person speaking or reporting the incident +- DO NOT extract victims, witnesses, or members of the public +- Example: "Officer Smith reporting... victims are John and Jane" → extract "Smith" +""" + elif any(w in field_lower for w in ['victim', 'injured', 'affected', 'casualty', 'patient']): + role_guidance = f""" +ROLE: Extract VICTIM/AFFECTED PERSON(S) +- Focus on people who experienced harm +- Ignore officers, dispatchers, and witnesses +{'- Return ALL names separated by ";"' if is_plural else '- Return the FIRST/PRIMARY victim'} +""" + elif any(w in field_lower for w in ['location', 'address', 'street', 'place', 'where']): + role_guidance = """ +ROLE: Extract LOCATION/ADDRESS +- Extract WHERE the incident occurred +- Return only the incident location, not other addresses mentioned +""" + elif any(w in field_lower for w in ['date', 'time', 'when', 'occurred', 'reported']): + role_guidance = """ +ROLE: Extract DATE/TIME +- Extract WHEN the incident occurred +- Return in the format it appears in the text +""" + elif any(w in field_lower for w in ['phone', 'number', 'contact', 'tel']): + role_guidance = "ROLE: Extract PHONE NUMBER — return exactly as it appears in text" + elif any(w in field_lower for w in ['email', 'mail']): + role_guidance = "ROLE: Extract EMAIL ADDRESS" + elif any(w in field_lower for w in ['department', 'unit', 'division']): + role_guidance = "ROLE: Extract DEPARTMENT/UNIT name" + elif any(w in field_lower for w in ['title', 'job', 'role', 'rank', 'position']): + role_guidance = "ROLE: Extract JOB TITLE or RANK" + elif any(w in field_lower for w in ['id', 'badge', 'identifier']): + role_guidance = "ROLE: Extract ID or BADGE NUMBER" + elif any(w in field_lower for w in ['description', 'incident', 'detail', 'nature', 'summary']): + role_guidance = "ROLE: Extract a brief INCIDENT DESCRIPTION" + else: + role_guidance = f""" +ROLE: Generic extraction for field "{current_field}" +{'- Return MULTIPLE values separated by ";" if applicable' if is_plural else '- Return the PRIMARY matching value'} +""" + + prompt = f""" +SYSTEM: You are extracting specific information from an incident report transcript. + +FIELD TO EXTRACT: {current_field} +{'[SINGULAR - Extract ONE value]' if not is_plural else '[PLURAL - Extract MULTIPLE values separated by semicolon]'} + +EXTRACTION RULES: +{role_guidance} + +CRITICAL RULES: +1. Read the ENTIRE text before answering +2. Extract ONLY what belongs to this specific field +3. Return values exactly as they appear in the text +4. If not found, return: -1 + +TRANSCRIPT: +{self._transcript_text} + +ANSWER: Return ONLY the extracted value(s), nothing else.""" + + return prompt + + async def async_main_loop(self): + """ + Async batch Ollama call — extracts ALL fields in one request. + Prevents blocking the FastAPI event loop during high-latency LLM calls. + """ + import httpx + ollama_host = os.getenv("OLLAMA_HOST", "http://localhost:11434").rstrip("/") + ollama_url = f"{ollama_host}/api/generate" + + if isinstance(self._target_fields, dict): + field_keys = list(self._target_fields.keys()) + field_names = list(self._target_fields.values()) + else: + field_keys = list(self._target_fields) + field_names = list(self._target_fields) + + field_count = len(field_keys) + print(f"[LOG] Starting async batch extraction for {field_count} field(s)...") + prompt = self.build_batch_prompt() + payload = {"model": "mistral", "prompt": prompt, "stream": False, "format": "json"} + + _start = time.time() + try: + timeout = int(os.getenv("OLLAMA_TIMEOUT", "300")) + async with httpx.AsyncClient() as client: + response = await client.post(ollama_url, json=payload, timeout=timeout) response.raise_for_status() - except requests.exceptions.ConnectionError: - raise ConnectionError( - f"Could not connect to Ollama at {ollama_url}. " - "Please ensure Ollama is running and accessible." - ) - except requests.exceptions.HTTPError as e: - raise RuntimeError(f"Ollama returned an error: {e}") - - # parse response - json_data = response.json() - parsed_response = json_data["response"] - # print(parsed_response) - self.add_response_to_json(field, parsed_response) - - print("----------------------------------") - print("\t[LOG] Resulting JSON created from the input text:") - print(json.dumps(self._json, indent=2)) - print("--------- extracted data ---------") + + _elapsed = time.time() - _start + print(f"[LOG] Ollama responded in {_elapsed:.2f}s") + raw = response.json()["response"].strip() + raw = raw.replace("```json", "").replace("```", "").strip() + + try: + extracted = json.loads(raw) + + # 1. First extract explicit keys mapped from templates + for key in field_keys: + val = extracted.get(key) + self._json[key] = val if val and str(val).lower() not in ("null", "none", "") else None + + # 2. Fully Dynamic Schema-less Extension: + # Accept EVERY OTHER valid key the LLM invented! + for key, val in extracted.items(): + if key not in field_keys: + if val and str(val).lower() not in ("null", "none", ""): + self._json[key] = val + + print("\t[LOG] Batch extraction successful.") + except json.JSONDecodeError: + print("\t[WARN] Batch JSON parse failed — falling back to per-field extraction") + # Fallback to sync for now or keep as is — usually batch works + self._json = {} + + except Exception as e: + print(f"[ERROR] Ollama request failed: {e}") + raise ConnectionError(f"Ollama connection failed: {e}") return self + def _fallback_per_field(self, ollama_url: str, field_keys: list): + """ + Legacy per-field extraction — used only when batch JSON parse fails. + """ + print("\t[LOG] Running fallback per-field extraction...") + + total = len(field_keys) + for i, field in enumerate(field_keys, 1): + print(f"[LOG] Extracting field {i}/{total} -> '{field}'") + if isinstance(self._target_fields, dict): + label = self._target_fields.get(field, field) + if not label or label == field: + label = field + else: + label = field + + prompt = self.build_prompt(label) + payload = {"model": "mistral", "prompt": prompt, "stream": False} + + try: + response = requests.post(ollama_url, json=payload) + response.raise_for_status() + parsed_response = response.json()["response"] + self.add_response_to_json(field, parsed_response) + except Exception as e: + print(f"\t[WARN] Failed to extract field '{field}': {e}") + self._json[field] = None + def add_response_to_json(self, field, value): """ - this method adds the following value under the specified field, - or under a new field if the field doesn't exist, to the json dict + Add extracted value under field name. + Handles plural (semicolon-separated) values. """ value = value.strip().replace('"', "") parsed_value = None @@ -94,42 +259,35 @@ def add_response_to_json(self, field, value): if value != "-1": parsed_value = value - if ";" in value: - parsed_value = self.handle_plural_values(value) + if parsed_value and ";" in parsed_value: + parsed_value = self.handle_plural_values(parsed_value) - if field in self._json.keys(): - self._json[field].append(parsed_value) + if field in self._json: + existing = self._json[field] + if isinstance(existing, list): + if isinstance(parsed_value, list): + existing.extend(parsed_value) + else: + existing.append(parsed_value) + else: + self._json[field] = [existing, parsed_value] else: self._json[field] = parsed_value - return - def handle_plural_values(self, plural_value): """ - This method handles plural values. - Takes in strings of the form 'value1; value2; value3; ...; valueN' - returns a list with the respective values -> [value1, value2, value3, ..., valueN] + Split semicolon-separated values into a list. + "Mark Smith; Jane Doe" → ["Mark Smith", "Jane Doe"] """ if ";" not in plural_value: raise ValueError( f"Value is not plural, doesn't have ; separator, Value: {plural_value}" ) - print( - f"\t[LOG]: Formating plural values for JSON, [For input {plural_value}]..." - ) - values = plural_value.split(";") - - # Remove trailing leading whitespace - for i in range(len(values)): - current = i + 1 - if current < len(values): - clean_value = values[current].lstrip() - values[current] = clean_value - + print(f"\t[LOG]: Formatting plural values for JSON, [For input {plural_value}]...") + values = [v.strip() for v in plural_value.split(";") if v.strip()] print(f"\t[LOG]: Resulting formatted list of values: {values}") - return values def get_data(self): - return self._json + return self._json \ No newline at end of file diff --git a/src/main.py b/src/main.py index 5bb632b..e07578b 100644 --- a/src/main.py +++ b/src/main.py @@ -1,5 +1,6 @@ import os # from backend import Fill +from typing import Union from commonforms import prepare_form from pypdf import PdfReader from controller import Controller diff --git a/src/transcriber.py b/src/transcriber.py new file mode 100644 index 0000000..ab0874f --- /dev/null +++ b/src/transcriber.py @@ -0,0 +1,58 @@ +import os +import tempfile +from pathlib import Path + + +def transcribe_audio(file_bytes: bytes, filename: str, language: str = None) -> dict: + """ + Transcribe audio file using faster-whisper. + + Args: + file_bytes: Raw audio file bytes + filename: Original filename (used to detect format) + language: Optional language code (e.g. 'en', 'fr'). None = auto-detect. + + Returns: + dict with 'transcript', 'language', 'duration' + + Supports: mp3, mp4, wav, m4a, ogg, webm (anything ffmpeg handles) + CPU-only — no GPU required. ~4x faster than openai-whisper, 3x less RAM. + """ + try: + from faster_whisper import WhisperModel + except ImportError: + raise RuntimeError( + "faster-whisper not installed. Run: pip install faster-whisper" + ) + + # Write bytes to temp file — faster-whisper needs a file path + suffix = Path(filename).suffix or ".wav" + with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp: + tmp.write(file_bytes) + tmp_path = tmp.name + + try: + # Use tiny model by default — fast, CPU-friendly, good accuracy + # Model downloads once (~75MB) to ~/.cache/huggingface/ + model_size = os.getenv("WHISPER_MODEL", "tiny") + model = WhisperModel(model_size, device="cpu", compute_type="int8") + + segments, info = model.transcribe( + tmp_path, + language=language, + beam_size=5, + vad_filter=True, # skip silent sections + vad_parameters=dict(min_silence_duration_ms=500) + ) + + transcript = " ".join(segment.text.strip() for segment in segments) + + return { + "transcript": transcript.strip(), + "language": info.language, + "language_probability": round(info.language_probability, 2), + "duration": round(info.duration, 1) + } + + finally: + os.unlink(tmp_path) \ No newline at end of file diff --git a/tests/conftest.py b/tests/conftest.py index 7cb4db3..82e4297 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -1,9 +1,8 @@ from fastapi.testclient import TestClient -from sqlmodel import SQLModel, create_engine, Session +from sqlmodel import SQLModel, create_engine, Session, delete from sqlalchemy.pool import StaticPool import pytest - from api.main import app from api.deps import get_db from api.db.models import Template, FormSubmission @@ -34,6 +33,46 @@ def create_test_db(): SQLModel.metadata.drop_all(engine) +@pytest.fixture(autouse=True) +def clean_db(): + """Wipe all tables before each test — prevents data leaking between tests.""" + with Session(engine) as session: + session.exec(delete(FormSubmission)) + session.exec(delete(Template)) + session.commit() + yield + + +@pytest.fixture +def db_session(): + """Provide a DB session for tests that need to insert data directly.""" + with Session(engine) as session: + yield session + + @pytest.fixture def client(): return TestClient(app) + + +@pytest.fixture +def tmp_pdf(tmp_path): + """ + Creates a real minimal PDF file on disk for tests. + Needed because forms.py validates pdf_path exists before calling Ollama. + """ + pdf_file = tmp_path / "test_form.pdf" + pdf_file.write_bytes( + b"%PDF-1.4\n" + b"1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n" + b"2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n" + b"3 0 obj\n<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] >>\nendobj\n" + b"xref\n0 4\n" + b"0000000000 65535 f\n" + b"0000000009 00000 n\n" + b"0000000058 00000 n\n" + b"0000000115 00000 n\n" + b"trailer\n<< /Size 4 /Root 1 0 R >>\n" + b"startxref\n190\n%%EOF\n" + ) + return str(pdf_file) \ No newline at end of file diff --git a/tests/test_filler.py b/tests/test_filler.py new file mode 100644 index 0000000..27f8c77 --- /dev/null +++ b/tests/test_filler.py @@ -0,0 +1,110 @@ +import pytest +from unittest.mock import MagicMock +from pdfrw import PdfName +from src.filler import _resolve_checkbox_value, _resolve_radio_kid, _get_field_type + + +class TestResolveCheckboxValue: + + def _make_annot(self, ap_keys=None): + annot = MagicMock() + if ap_keys: + annot.AP.N.keys.return_value = [f"/{k}" for k in ap_keys] + else: + annot.AP = None + return annot + + def test_yes_string_returns_pdf_yes(self): + annot = self._make_annot(["Yes", "Off"]) + assert str(_resolve_checkbox_value("yes", annot)) == "/Yes" + + def test_true_string_returns_checked(self): + annot = self._make_annot(["Yes", "Off"]) + assert str(_resolve_checkbox_value("true", annot)) != "/Off" + + def test_no_string_returns_off(self): + annot = self._make_annot(["Yes", "Off"]) + assert str(_resolve_checkbox_value("no", annot)) == "/Off" + + def test_false_string_returns_off(self): + annot = self._make_annot() + assert str(_resolve_checkbox_value("false", annot)) == "/Off" + + def test_empty_string_returns_off(self): + annot = self._make_annot() + assert str(_resolve_checkbox_value("", annot)) == "/Off" + + def test_no_ap_falls_back_to_yes(self): + annot = self._make_annot() + assert str(_resolve_checkbox_value("yes", annot)) == "/Yes" + + def test_custom_on_value_from_ap(self): + annot = self._make_annot(["On", "Off"]) + assert str(_resolve_checkbox_value("yes", annot)) == "/On" + + def test_x_means_checked(self): + annot = self._make_annot(["Yes", "Off"]) + assert str(_resolve_checkbox_value("x", annot)) != "/Off" + + def test_none_value_returns_off(self): + annot = self._make_annot() + assert str(_resolve_checkbox_value("none", annot)) == "/Off" + + +class TestResolveRadioKid: + + def _make_annot(self, ap_keys=None, opt_list=None): + annot = MagicMock() + if ap_keys: + annot.AP.N.keys.return_value = [f"/{k}" for k in ap_keys] + else: + annot.AP = None + if opt_list: + annot.Parent.Opt = [f"({o})" for o in opt_list] + else: + annot.Parent = None + return annot + + def test_selected_returns_option_value(self): + """kid_index=0, raw='male', opt=['Male','Female'] → /Male""" + annot = self._make_annot(ap_keys=["Male", "Off"], opt_list=["Male", "Female"]) + result = _resolve_radio_kid("male", 0, annot) + assert str(result) == "/Male" + + def test_unselected_returns_off(self): + """kid_index=0 is Male but raw='female' → /Off""" + annot = self._make_annot(ap_keys=["Male", "Off"], opt_list=["Male", "Female"]) + result = _resolve_radio_kid("female", 0, annot) + assert str(result) == "/Off" + + def test_no_parent_returns_off(self): + """No parent opt list → cannot determine selection → /Off""" + annot = self._make_annot() + result = _resolve_radio_kid("yes", 0, annot) + assert str(result) == "/Off" + + +class TestGetFieldType: + + def _make_annot(self, ft, ff=0): + annot = MagicMock() + annot.FT = f"/{ft}" + annot.Ff = str(ff) + return annot + + def test_text_field(self): + assert _get_field_type(self._make_annot("Tx")) == "text" + + def test_checkbox_field(self): + assert _get_field_type(self._make_annot("Btn", ff=0)) == "checkbox" + + def test_radio_field(self): + assert _get_field_type(self._make_annot("Btn", ff=1 << 15)) == "radio" + + def test_unknown_field_type(self): + assert _get_field_type(self._make_annot("Sig")) == "other" + + def test_no_ft_returns_other(self): + annot = MagicMock() + annot.FT = None + assert _get_field_type(annot) == "other" \ No newline at end of file diff --git a/tests/test_forms.py b/tests/test_forms.py index 8f432bf..f55eedb 100644 --- a/tests/test_forms.py +++ b/tests/test_forms.py @@ -1,25 +1,120 @@ -def test_submit_form(client): - pass - # First create a template - # form_payload = { - # "template_id": 3, - # "input_text": "Hi. The employee's name is John Doe. His job title is managing director. His department supervisor is Jane Doe. His phone number is 123456. His email is jdoe@ucsc.edu. The signature is , and the date is 01/02/2005", - # } - - # template_res = client.post("/templates/", json=template_payload) - # template_id = template_res.json()["id"] - - # # Submit a form - # form_payload = { - # "template_id": template_id, - # "data": {"rating": 5, "comment": "Great service"}, - # } - - # response = client.post("/forms/", json=form_payload) - - # assert response.status_code == 200 - - # data = response.json() - # assert data["id"] is not None - # assert data["template_id"] == template_id - # assert data["data"] == form_payload["data"] +""" +Tests for /forms endpoints. +Closes #165, #205, #163 +""" + +import pytest +import os +from unittest.mock import patch +from api.db.models import Template, FormSubmission +from datetime import datetime + + +# ── helpers ────────────────────────────────────────────────────────────────── + +def make_template(db_session, pdf_path="/tmp/test.pdf"): + t = Template( + name="Test Form", + fields={"JobTitle": "Job Title"}, + pdf_path=pdf_path, + created_at=datetime.utcnow(), + ) + db_session.add(t) + db_session.commit() + db_session.refresh(t) + return t.id + + +def make_submission(db_session, template_id, output_path="/tmp/filled.pdf"): + s = FormSubmission( + template_id=template_id, + input_text="John Smith is a firefighter.", + output_pdf_path=output_path, + created_at=datetime.utcnow(), + ) + db_session.add(s) + db_session.commit() + db_session.refresh(s) + return s.id + + +# ── POST /forms/fill ────────────────────────────────────────────────────────── + +class TestFillForm: + + def test_fill_form_template_not_found(self, client): + """Returns 404 when template_id does not exist.""" + response = client.post("/forms/fill", json={ + "template_id": 999999, + "input_text": "John Smith is a firefighter.", + }) + assert response.status_code == 404 + + def test_fill_form_missing_fields_returns_422(self, client): + """Returns 422 when required fields are missing.""" + response = client.post("/forms/fill", json={ + "template_id": 1, + }) + assert response.status_code == 422 + + def test_fill_form_ollama_down_returns_503(self, client, db_session, tmp_pdf): + """Returns 503 when Ollama is not reachable.""" + # Use tmp_pdf so our pdf_path validation passes before hitting Ollama + template_id = make_template(db_session, pdf_path=tmp_pdf) + + with patch("src.controller.Controller.fill_form", + side_effect=ConnectionError("Ollama not running")): + response = client.post("/forms/fill", json={ + "template_id": template_id, + "input_text": "John Smith is a firefighter.", + }) + + assert response.status_code == 503 + + def test_fill_form_pdf_not_on_disk_returns_404(self, client, db_session): + """Returns 404 when template PDF path does not exist on disk.""" + template_id = make_template(db_session, pdf_path="/nonexistent/path.pdf") + + response = client.post("/forms/fill", json={ + "template_id": template_id, + "input_text": "John Smith is a firefighter.", + }) + + assert response.status_code == 404 + + +# ── GET /forms/{submission_id} ──────────────────────────────────────────────── + +class TestGetSubmission: + + def test_get_submission_not_found(self, client): + """Returns 404 for non-existent submission ID.""" + response = client.get("/forms/999999") + assert response.status_code == 404 + + def test_get_submission_invalid_id(self, client): + """Returns 422 for non-integer submission ID.""" + response = client.get("/forms/not-an-id") + assert response.status_code == 422 + + +# ── GET /forms/download/{submission_id} ────────────────────────────────────── + +class TestDownloadSubmission: + + def test_download_not_found_submission(self, client): + """Returns 404 when submission does not exist.""" + response = client.get("/forms/download/999999") + assert response.status_code == 404 + + def test_download_file_missing_on_disk(self, client, db_session): + """Returns 404 when submission exists but PDF missing on disk.""" + template_id = make_template(db_session) + submission_id = make_submission( + db_session, template_id, "/nonexistent/filled.pdf" + ) + + with patch("os.path.exists", return_value=False): + response = client.get(f"/forms/download/{submission_id}") + + assert response.status_code == 404 \ No newline at end of file diff --git a/tests/test_incidents.py b/tests/test_incidents.py new file mode 100644 index 0000000..4284d75 --- /dev/null +++ b/tests/test_incidents.py @@ -0,0 +1,276 @@ +""" +Tests for the Master Incident Data Lake — PR #1. + +These tests cover: +- Creating a new incident record via POST /incidents/extract +- Retrieving an incident via GET /incidents/{id} +- Collaborative Consensus Merge (multi-officer append) +- 404 handling for unknown incidents / templates +- PDF generation from stored Data Lake record + +The LLM (Ollama/Mistral) is mocked in all tests — no running +Ollama instance is required. +""" + +import json +import pytest +from unittest.mock import patch, AsyncMock + +from fastapi.testclient import TestClient +from sqlmodel import SQLModel, Session, create_engine, delete +from sqlalchemy.pool import StaticPool + +from api.main import app +from api.deps import get_db +from api.db.models import Template, FormSubmission, IncidentMasterData +from api.db.repositories import ( + create_incident, + get_incident, + update_incident_json, +) + +# ── In-memory test database ──────────────────────────────────────────── + +TEST_DB_URL = "sqlite://" +engine = create_engine( + TEST_DB_URL, + connect_args={"check_same_thread": False}, + poolclass=StaticPool, +) + + +def override_get_db(): + with Session(engine) as session: + yield session + + +app.dependency_overrides[get_db] = override_get_db + + +@pytest.fixture(scope="session", autouse=True) +def create_test_db(): + SQLModel.metadata.create_all(engine) + yield + SQLModel.metadata.drop_all(engine) + + +@pytest.fixture(autouse=True) +def clean_db(): + """Wipe all tables before each test — prevents leakage between tests.""" + with Session(engine) as session: + session.exec(delete(FormSubmission)) + session.exec(delete(IncidentMasterData)) + session.exec(delete(Template)) + session.commit() + yield + + +@pytest.fixture +def db_session(): + with Session(engine) as session: + yield session + + +@pytest.fixture +def client(): + return TestClient(app) + + +# ── Mock LLM response ───────────────────────────────────────────────── + +MOCK_EXTRACTED = { + "OfficerName": "John Smith", + "BadgeNumber": "EMP-001", + "Location": "742 Evergreen Terrace", + "IncidentType": "Structure Fire", +} + + +def make_mock_llm(): + """Returns a mock LLM object whose async_main_loop does nothing and get_data returns mock data.""" + mock = AsyncMock() + mock.async_main_loop = AsyncMock(return_value=None) + mock.get_data = lambda: MOCK_EXTRACTED + return mock + + +# ── Unit Tests: Consensus Merge (no HTTP) ───────────────────────────── + +class TestConsensusRepositoryLogic: + + def test_create_incident_persists(self, db_session): + """Creating an incident stores it in the database.""" + incident = IncidentMasterData( + incident_id="INC-UNIT-001", + master_json=json.dumps({"OfficerName": "Alice"}), + transcript_text="Officer Alice on scene.", + ) + saved = create_incident(db_session, incident) + assert saved.id is not None + assert saved.incident_id == "INC-UNIT-001" + + def test_get_incident_retrieves_correct_record(self, db_session): + """get_incident returns the correct record by incident_id.""" + incident = IncidentMasterData( + incident_id="INC-UNIT-002", + master_json=json.dumps({"OfficerName": "Bob"}), + transcript_text="Officer Bob reporting.", + ) + create_incident(db_session, incident) + retrieved = get_incident(db_session, "INC-UNIT-002") + assert retrieved is not None + assert retrieved.incident_id == "INC-UNIT-002" + + def test_get_incident_returns_none_for_unknown(self, db_session): + """get_incident returns None when incident does not exist.""" + result = get_incident(db_session, "INC-DOES-NOT-EXIST") + assert result is None + + def test_consensus_merge_does_not_overwrite_with_null(self, db_session): + """Smart merge: null/None values do NOT overwrite existing valid data.""" + incident = IncidentMasterData( + incident_id="INC-MERGE-001", + master_json=json.dumps({"OfficerName": "Alice", "BadgeNumber": "EMP-001"}), + transcript_text="First report.", + ) + create_incident(db_session, incident) + + # Second officer sends None for OfficerName — should NOT overwrite + update_incident_json( + db_session, + "INC-MERGE-001", + {"OfficerName": None, "Location": "742 Evergreen Terrace"}, + new_transcript="Second report.", + ) + + updated = get_incident(db_session, "INC-MERGE-001") + result = json.loads(updated.master_json) + assert result["OfficerName"] == "Alice" # protected + assert result["Location"] == "742 Evergreen Terrace" # new field added + + def test_consensus_merge_appends_notes_field(self, db_session): + """Smart merge: long-form text fields (Notes) append with [UPDATE] tag.""" + incident = IncidentMasterData( + incident_id="INC-MERGE-002", + master_json=json.dumps({"Notes": "Fire on ground floor."}), + transcript_text="Initial note.", + ) + create_incident(db_session, incident) + + update_incident_json( + db_session, + "INC-MERGE-002", + {"Notes": "Victim evacuated safely."}, + new_transcript="Second note.", + ) + + updated = get_incident(db_session, "INC-MERGE-002") + result = json.loads(updated.master_json) + assert "Fire on ground floor." in result["Notes"] + assert "[UPDATE]" in result["Notes"] + assert "Victim evacuated safely." in result["Notes"] + + def test_consensus_merge_overwrites_short_fields_with_new_data(self, db_session): + """Regular (non-notes) fields with real new values DO get updated.""" + incident = IncidentMasterData( + incident_id="INC-MERGE-003", + master_json=json.dumps({"Location": "Old Address"}), + transcript_text="Initial.", + ) + create_incident(db_session, incident) + + update_incident_json( + db_session, + "INC-MERGE-003", + {"Location": "New Corrected Address"}, + new_transcript="Correction.", + ) + + updated = get_incident(db_session, "INC-MERGE-003") + result = json.loads(updated.master_json) + assert result["Location"] == "New Corrected Address" + + +# ── Integration Tests: API Endpoints ────────────────────────────────── + +class TestDataLakeEndpoints: + + def test_extract_creates_new_incident(self, client): + """POST /incidents/extract creates a new incident record.""" + with patch("api.routes.incidents.LLM", return_value=make_mock_llm()): + response = client.post( + "/incidents/extract", + params={ + "input_text": "Officer John Smith EMP-001 structure fire 742 Evergreen Terrace.", + "incident_id": "INC-E2E-001", + }, + ) + assert response.status_code == 200 + data = response.json() + assert data["incident_id"] == "INC-E2E-001" + assert data["status"] == "created" + + def test_extract_merges_into_existing_incident(self, client): + """POST /incidents/extract with same ID returns status 'merged'.""" + with patch("api.routes.incidents.LLM", return_value=make_mock_llm()): + client.post( + "/incidents/extract", + params={"input_text": "First officer report.", "incident_id": "INC-E2E-002"}, + ) + response = client.post( + "/incidents/extract", + params={"input_text": "Second officer adding location.", "incident_id": "INC-E2E-002"}, + ) + assert response.status_code == 200 + assert response.json()["status"] == "merged" + + def test_get_incident_returns_stored_data(self, client, db_session): + """GET /incidents/{id} returns the stored master JSON.""" + incident = IncidentMasterData( + incident_id="INC-GET-001", + master_json=json.dumps({"OfficerName": "Alice"}), + transcript_text="Officer Alice.", + ) + create_incident(db_session, incident) + + response = client.get("/incidents/INC-GET-001") + assert response.status_code == 200 + data = response.json() + assert data["incident_id"] == "INC-GET-001" + assert data["master_json"]["OfficerName"] == "Alice" + + def test_get_nonexistent_incident_returns_404(self, client): + """GET /incidents/{id} returns 404 for unknown ID.""" + response = client.get("/incidents/INC-GHOST-999") + assert response.status_code == 404 + + def test_generate_returns_404_for_missing_incident(self, client): + """POST /incidents/{id}/generate/{template_id} returns 404 when incident missing.""" + response = client.post("/incidents/INC-MISSING/generate/1") + assert response.status_code == 404 + + def test_generate_returns_404_for_missing_template(self, client, db_session): + """POST /incidents/{id}/generate/{template_id} returns 404 when template missing.""" + incident = IncidentMasterData( + incident_id="INC-GEN-001", + master_json=json.dumps({"OfficerName": "Alice"}), + transcript_text="Officer Alice.", + ) + create_incident(db_session, incident) + response = client.post("/incidents/INC-GEN-001/generate/99999") + assert response.status_code == 404 + + def test_list_all_incidents(self, client, db_session): + """GET /incidents returns a list of all stored incidents.""" + for i in range(3): + create_incident( + db_session, + IncidentMasterData( + incident_id=f"INC-LIST-00{i}", + master_json=json.dumps({}), + transcript_text="test", + ), + ) + response = client.get("/incidents") + assert response.status_code == 200 + assert len(response.json()) >= 3 \ No newline at end of file diff --git a/tests/test_llm.py b/tests/test_llm.py new file mode 100644 index 0000000..cfe483b --- /dev/null +++ b/tests/test_llm.py @@ -0,0 +1,278 @@ +""" +Unit tests for src/llm.py — LLM class. + +Closes: #186 (Unit tests for LLM class methods) +Covers: batch prompt, per-field prompt, add_response_to_json, + handle_plural_values, type_check_all, main_loop (mocked) +""" + +import json +import pytest +from unittest.mock import patch, MagicMock +from src.llm import LLM + + +# ── Fixtures ───────────────────────────────────────────────────────────────── + +@pytest.fixture +def dict_fields(): + """Realistic dict fields: {internal_name: human_label}""" + return { + "NAME/SID": "Employee Or Student Name", + "JobTitle": "Job Title", + "Department": "Department", + "Phone Number": "Phone Number", + "email": "Email", + } + +@pytest.fixture +def list_fields(): + """Legacy list fields: [internal_name, ...]""" + return ["officer_name", "location", "incident_date"] + +@pytest.fixture +def transcript(): + return ( + "Employee name is John Smith. Employee ID is EMP-2024-789. " + "Job title is Firefighter Paramedic. Department is Emergency Medical Services. " + "Phone number is 916-555-0147." + ) + +@pytest.fixture +def llm_dict(dict_fields, transcript): + return LLM(transcript_text=transcript, target_fields=dict_fields) + +@pytest.fixture +def llm_list(list_fields, transcript): + return LLM(transcript_text=transcript, target_fields=list_fields) + + +# ── type_check_all ──────────────────────────────────────────────────────────── + +class TestTypeCheckAll: + + def test_raises_on_non_string_transcript(self, dict_fields): + llm = LLM(transcript_text=12345, target_fields=dict_fields) + with pytest.raises(TypeError, match="Transcript must be text"): + llm.type_check_all() + + def test_raises_on_none_transcript(self, dict_fields): + llm = LLM(transcript_text=None, target_fields=dict_fields) + with pytest.raises(TypeError): + llm.type_check_all() + + def test_raises_on_invalid_fields_type(self, transcript): + llm = LLM(transcript_text=transcript, target_fields="not_a_list_or_dict") + with pytest.raises(TypeError, match="list or dict"): + llm.type_check_all() + + def test_passes_with_dict_fields(self, llm_dict): + # Should not raise + llm_dict.type_check_all() + + def test_passes_with_list_fields(self, llm_list): + # Should not raise + llm_list.type_check_all() + + +# ── build_batch_prompt ──────────────────────────────────────────────────────── + +class TestBuildBatchPrompt: + + def test_contains_all_field_keys(self, llm_dict, dict_fields): + prompt = llm_dict.build_batch_prompt() + for key in dict_fields.keys(): + assert key in prompt, f"Field key '{key}' missing from batch prompt" + + def test_contains_human_labels(self, llm_dict, dict_fields): + prompt = llm_dict.build_batch_prompt() + for label in dict_fields.values(): + assert label in prompt, f"Label '{label}' missing from batch prompt" + + def test_contains_transcript(self, llm_dict, transcript): + prompt = llm_dict.build_batch_prompt() + assert transcript in prompt + + def test_contains_json_instruction(self, llm_dict): + prompt = llm_dict.build_batch_prompt() + assert "JSON" in prompt + + def test_list_fields_batch_prompt(self, llm_list, list_fields): + prompt = llm_list.build_batch_prompt() + for field in list_fields: + assert field in prompt + + def test_labels_used_as_comments(self, llm_dict): + """Human labels should appear after // in the prompt""" + prompt = llm_dict.build_batch_prompt() + assert "//" in prompt + + +# ── build_prompt (legacy per-field) ────────────────────────────────────────── + +class TestBuildPrompt: + + def test_officer_field_gets_officer_guidance(self, llm_dict): + prompt = llm_dict.build_prompt("officer_name") + assert "OFFICER" in prompt.upper() or "EMPLOYEE" in prompt.upper() + + def test_location_field_gets_location_guidance(self, llm_dict): + prompt = llm_dict.build_prompt("incident_location") + assert "LOCATION" in prompt.upper() or "ADDRESS" in prompt.upper() + + def test_victim_field_gets_victim_guidance(self, llm_dict): + prompt = llm_dict.build_prompt("victim_name") + assert "VICTIM" in prompt.upper() + + def test_phone_field_gets_phone_guidance(self, llm_dict): + prompt = llm_dict.build_prompt("Phone Number") + assert "PHONE" in prompt.upper() + + def test_prompt_contains_transcript(self, llm_dict, transcript): + prompt = llm_dict.build_prompt("some_field") + assert transcript in prompt + + def test_generic_field_still_builds_prompt(self, llm_dict): + prompt = llm_dict.build_prompt("textbox_0_0") + assert len(prompt) > 50 + + +# ── handle_plural_values ────────────────────────────────────────────────────── + +class TestHandlePluralValues: + + def test_splits_on_semicolon(self, llm_dict): + result = llm_dict.handle_plural_values("Mark Smith;Jane Doe") + assert "Mark Smith" in result + assert "Jane Doe" in result + + def test_strips_whitespace(self, llm_dict): + result = llm_dict.handle_plural_values("Mark Smith; Jane Doe; Bob") + assert all(v == v.strip() for v in result) + + def test_returns_list(self, llm_dict): + result = llm_dict.handle_plural_values("A;B;C") + assert isinstance(result, list) + + def test_raises_without_semicolon(self, llm_dict): + with pytest.raises(ValueError, match="separator"): + llm_dict.handle_plural_values("no semicolon here") + + def test_three_values(self, llm_dict): + result = llm_dict.handle_plural_values("Alice;Bob;Charlie") + assert len(result) == 3 + + +# ── add_response_to_json ────────────────────────────────────────────────────── + +class TestAddResponseToJson: + + def test_stores_value_under_field(self, llm_dict): + llm_dict.add_response_to_json("NAME/SID", "John Smith") + assert llm_dict._json["NAME/SID"] == "John Smith" + + def test_ignores_minus_one(self, llm_dict): + llm_dict.add_response_to_json("email", "-1") + assert llm_dict._json["email"] is None + + def test_strips_quotes(self, llm_dict): + llm_dict.add_response_to_json("JobTitle", '"Firefighter"') + assert llm_dict._json["JobTitle"] == "Firefighter" + + def test_strips_whitespace(self, llm_dict): + llm_dict.add_response_to_json("Department", " EMS ") + assert llm_dict._json["Department"] == "EMS" + + def test_plural_value_becomes_list(self, llm_dict): + llm_dict.add_response_to_json("victims", "Mark Smith;Jane Doe") + assert isinstance(llm_dict._json["victims"], list) + + def test_existing_field_becomes_list(self, llm_dict): + """Adding to existing field should not overwrite silently.""" + llm_dict._json["NAME/SID"] = "John" + llm_dict.add_response_to_json("NAME/SID", "Jane") + assert isinstance(llm_dict._json["NAME/SID"], list) + + +# ── get_data ────────────────────────────────────────────────────────────────── + +class TestGetData: + + def test_returns_dict(self, llm_dict): + assert isinstance(llm_dict.get_data(), dict) + + def test_returns_same_reference_as_internal_json(self, llm_dict): + llm_dict._json["test_key"] = "test_value" + assert llm_dict.get_data()["test_key"] == "test_value" + + +# ── main_loop (mocked Ollama) ───────────────────────────────────────────────── + +class TestMainLoop: + + def _mock_response(self, json_body: dict): + """Build a mock requests.Response returning a valid Mistral JSON reply.""" + mock_resp = MagicMock() + mock_resp.raise_for_status = MagicMock() + mock_resp.json.return_value = { + "response": json.dumps(json_body) + } + return mock_resp + + def test_batch_success_fills_all_fields(self, llm_dict, dict_fields): + expected = { + "NAME/SID": "John Smith", + "JobTitle": "Firefighter Paramedic", + "Department": "Emergency Medical Services", + "Phone Number": "916-555-0147", + "email": None, + } + with patch("requests.post", return_value=self._mock_response(expected)): + llm_dict.main_loop() + + result = llm_dict.get_data() + assert result["NAME/SID"] == "John Smith" + assert result["JobTitle"] == "Firefighter Paramedic" + assert result["Department"] == "Emergency Medical Services" + assert result["Phone Number"] == "916-555-0147" + + def test_batch_makes_exactly_one_ollama_call(self, llm_dict, dict_fields): + """Core performance requirement — O(1) not O(N).""" + expected = {k: "value" for k in dict_fields.keys()} + with patch("requests.post", return_value=self._mock_response(expected)) as mock_post: + llm_dict.main_loop() + + assert mock_post.call_count == 1, ( + f"Expected 1 Ollama call, got {mock_post.call_count}. " + "main_loop() must use batch extraction, not per-field." + ) + + def test_fallback_on_invalid_json(self, llm_dict, dict_fields): + """If Mistral returns non-JSON, fallback per-field runs without crash.""" + bad_response = MagicMock() + bad_response.raise_for_status = MagicMock() + bad_response.json.return_value = {"response": "This is not JSON at all."} + + good_response = MagicMock() + good_response.raise_for_status = MagicMock() + good_response.json.return_value = {"response": "John Smith"} + + # First call returns bad JSON, rest return single values + with patch("requests.post", side_effect=[bad_response] + [good_response] * len(dict_fields)): + llm_dict.main_loop() # should not raise + + def test_connection_error_raises_connection_error(self, llm_dict): + import requests as req + with patch("requests.post", side_effect=req.exceptions.ConnectionError): + with pytest.raises(ConnectionError, match="Ollama"): + llm_dict.main_loop() + + def test_null_values_stored_as_none(self, llm_dict, dict_fields): + """Mistral returning null should be stored as None, not the string 'null'.""" + response_with_nulls = {k: None for k in dict_fields.keys()} + with patch("requests.post", return_value=self._mock_response(response_with_nulls)): + llm_dict.main_loop() + + result = llm_dict.get_data() + for key in dict_fields.keys(): + assert result[key] is None, f"Expected None for '{key}', got {result[key]!r}" diff --git a/tests/test_templates.py b/tests/test_templates.py index bbced2b..9b7cf8e 100644 --- a/tests/test_templates.py +++ b/tests/test_templates.py @@ -1,18 +1,126 @@ -def test_create_template(client): - payload = { - "name": "Template 1", - "pdf_path": "src/inputs/file.pdf", - "fields": { - "Employee's name": "string", - "Employee's job title": "string", - "Employee's department supervisor": "string", - "Employee's phone number": "string", - "Employee's email": "string", - "Signature": "string", - "Date": "string", - }, - } - - response = client.post("/templates/create", json=payload) - - assert response.status_code == 200 +""" +Tests for /templates endpoints. +Closes #162, #160, #163 +""" + +import io +import pytest +from unittest.mock import patch, MagicMock +from api.db.models import Template +from datetime import datetime + + +# ── POST /templates/create ──────────────────────────────────────────────────── + +class TestCreateTemplate: + + def test_create_template_success(self, client): + """Uploading a valid PDF returns 200 with template data.""" + pdf_bytes = ( + b"%PDF-1.4\n1 0 obj<>endobj\n" + b"2 0 obj<>endobj\n" + b"3 0 obj<>endobj\n" + b"xref\n0 4\n0000000000 65535 f\n" + b"trailer<>\nstartxref\n0\n%%EOF" + ) + + mock_fields = { + "JobTitle": {"/T": "JobTitle", "/FT": "/Tx"}, + "Department": {"/T": "Department", "/FT": "/Tx"}, + } + + with patch("commonforms.prepare_form"), \ + patch("pypdf.PdfReader") as mock_reader, \ + patch("shutil.copyfileobj"), \ + patch("builtins.open", MagicMock()), \ + patch("os.path.exists", return_value=True), \ + patch("os.remove"): + + mock_reader.return_value.get_fields.return_value = mock_fields + + response = client.post( + "/templates/create", + files={"file": ("form.pdf", io.BytesIO(pdf_bytes), "application/pdf")}, + data={"name": "Vaccine Form"}, + ) + + assert response.status_code == 200 + data = response.json() + assert data["name"] == "Vaccine Form" + assert "id" in data + assert "fields" in data + + def test_create_template_without_file_returns_422(self, client): + """Missing file field returns 422 Unprocessable Entity.""" + response = client.post( + "/templates/create", + data={"name": "No File"}, + ) + assert response.status_code == 422 + + def test_create_template_non_pdf_returns_400(self, client): + """Uploading a non-PDF returns 400.""" + with patch("shutil.copyfileobj"), \ + patch("builtins.open", MagicMock()): + response = client.post( + "/templates/create", + files={"file": ("data.csv", io.BytesIO(b"a,b,c"), "text/csv")}, + data={"name": "CSV attempt"}, + ) + assert response.status_code == 400 + + +# ── GET /templates ──────────────────────────────────────────────────────────── + +class TestListTemplates: + + def test_list_templates_returns_200(self, client): + """GET /templates returns 200.""" + response = client.get("/templates") + assert response.status_code == 200 + + def test_list_templates_returns_list(self, client): + """Response is always a list.""" + response = client.get("/templates") + assert isinstance(response.json(), list) + + def test_list_templates_empty_on_fresh_db(self, client): + """Fresh DB returns empty list.""" + response = client.get("/templates") + assert response.json() == [] + + def test_list_templates_pagination_accepted(self, client): + """Pagination params accepted without error.""" + response = client.get("/templates?limit=5&offset=0") + assert response.status_code == 200 + + +# ── GET /templates/{template_id} ────────────────────────────────────────────── + +class TestGetTemplate: + + def test_get_template_not_found(self, client): + """Returns 404 for non-existent ID.""" + response = client.get("/templates/999999") + assert response.status_code == 404 + + def test_get_template_invalid_id_type(self, client): + """Returns 422 for non-integer ID.""" + response = client.get("/templates/not-an-id") + assert response.status_code == 422 + + def test_get_template_by_id(self, client, db_session): + """Returns correct template for valid ID.""" + t = Template( + name="Cal Fire Form", + fields={"officer_name": "Officer Name"}, + pdf_path="/tmp/cal_fire.pdf", + created_at=datetime.utcnow(), + ) + db_session.add(t) + db_session.commit() + db_session.refresh(t) + + response = client.get(f"/templates/{t.id}") + assert response.status_code == 200 + assert response.json()["name"] == "Cal Fire Form"