Pre-Prod Task List — SOLVRO MCP (MVP Blockers/Improvements)

[W1ndrunn3rr]

Reference: Priority order for MVP launch

---

1. Persistent User & Session Storage

• Replace in-memory SessionManager with a real database (PostgreSQL or MongoDB as a new compose.stack.yml service)
• Migrate ConversationSession and Message models to SQLAlchemy / Motor async ORM
• Store: user_id, session_id, messages[], created_at, is_active, metadata
• Add user registration/auth (JWT or API key) — currently any user_id string is accepted with zero validation
• Migrate session_manager.py to async DB calls; drop the threading.Lock

2. Multi-threaded / Concurrent Data Pipeline

• Make the Prefect pipeline process documents in parallel
• Replace sequential for page in pages: loop in pipeline.py with asyncio.gather() or Prefect's task.submit() with thread/process pool
• Each page: extract → generate Cypher → populate runs concurrently (configurable concurrency limit)
• Add idempotency: track processed documents (hash → DB)
• Write integration test: mock 10 pages and verify concurrent write without schema reflection race

3. Google Drive as Data Source

• Replace Azure Blob with Google Drive in data_acquisition.py
• Authenticate via Google Drive API (service account JSON, secret-managed)
• List files from configured Drive folder (support PDF, DOCX, TXT)
• Download to temp dir, pass paths to OCR extraction
• Add GOOGLE_DRIVE_FOLDER_ID and GOOGLE_SERVICE_ACCOUNT_JSON to .env.example
• Handle pagination for large folders

4. Neo4j Graph State Snapshot (Data Dump)

• Export full graph after pipeline via CALL apoc.export.cypher.all() to .cypher dump (stored in cloud)
• On pipeline startup, check for dump and import if present, skipping LLM extraction
• Add just dump-graph and just restore-graph recipes
• Track pipeline_run metadata in Neo4j, only process new/changed files

5. Frontend Improvements

• Conversation naming (LLM-powered title after first reply; store in session.metadata["title"])
• Conversation list sidebar (titles, timestamp, rename/delete)
• Message streaming (SSE for /api/chat)
• Empty and error states (suggested questions, API error surfacing)

6. Component & Integration Tests

• Core logic test coverage is near zero; see recommended coverage targets for:
  • Guardrails, schema caching, cypher generation, pipeline execution
  • Full API/graph integration tests with real or mocked backends

7. (Extra) Query Caching & Rate Limiting

• Add semantic query cache (Redis + pgvector); check for near-duplicates before running LLM/graph
• Rate limiting on /api/chat (slowapi middleware), default 20 req/min per user
• Both prod features toggleable in .env

---

Blocking/effort tags:

• Graph dump/restore — Low effort, blocks prod (MVP must-have)
• Google Drive source — Medium effort, blocks prod
• Tests — Medium effort, deosn't blocks prod
• Persistent storage — High effort, blocks prod (sessions lost on restart!)
• Concurrent pipeline — Medium, not a blocker
• Frontend UX — Medium, not a blocker
• Cache/rate limit — Low, not a blocker

---

See full task list above for details. Prioritize by block/effort.

---