Skip to content

project.md

Latest commit

 

History

History
122 lines (90 loc) · 3.44 KB

project.md

File metadata and controls

122 lines (90 loc) · 3.44 KB

Context Engine

What It Is

Context Engine is an open-source structured context infrastructure for AI systems. It extracts structured facts from documents, organizes them into semantic models, and surfaces relationships as a living knowledge graph.

Core Problem

Company knowledge is scattered across Slack, Notion, Zoom transcripts, GitHub issues, meeting notes, and local documents. Traditional RAG gives you retrieval, but it doesn't answer:

  • Where did this answer come from?
  • Is this still true?
  • What changed?
  • How does pricing relate to features?

Context Engine's philosophy: facts should be structured, answers should cite sources, and relationships should be visible—not silently averaged away.

Architecture

Backend

Layer Technology
Framework FastAPI (async Python 3.12+)
ORM SQLAlchemy 2.0
Database SQLite (default), optional PostgreSQL + pgvector
LLM Gateway LiteLLM (OpenAI, Anthropic, etc.)
Embeddings LiteLLM or local HashingEmbedder (tests/offline)
Build Hatchling
CLI ctxe

Frontend

Layer Technology
Framework React 18
Build Tool Vite
Styling Tailwind CSS
Graph Viz Cytoscape.js
Routing React Router DOM v6

Infrastructure

  • Single Docker container (Python 3.12 slim + pre-built frontend)
  • SQLite default, PostgreSQL optional
  • No Redis, no Celery, no worker processes
  • Memory footprint: ~150MB RAM idle

Data Model

6 tables:

  • SourceDocument — Raw ingested content with metadata
  • Model — Product domain (Pricing, Features, Roadmap, Decisions)
  • Component — Atomic structured fact with confidence, status, embedding
  • Relationship — Typed edge between components
  • Connector — Catalogued or configured ingestion surface
  • SyncJob — Connector sync attempt and status record

Processing Pipeline

  1. Ingest document via API or CLI
  2. Extract structured facts (LLM or regex fallback)
  3. Create/update Model and Component records
  4. Batch embed new components
  5. Create Relationship edges from extracted links
  6. Query via semantic search + graph traversal

Product Workflows

3 frontend views:

  • Graph Explorer — Visual knowledge graph with models, components, and relationships
  • Ask (Query) — Natural language query with cited components and sources
  • Source Manager — Upload, browse, and inspect source documents

Operations

# One-command start
ctxe serve

# Ingest
ctxe ingest ./docs/

# Query
ctxe query "What is our pricing?"

# MCP server
ctxe mcp

Design Philosophy

  • Graph-first — Relationships are first-class, not afterthoughts
  • Self-host first — Single container, zero external dependencies by default
  • Minimal but impactful — 6 tables, focused views, 1 process
  • Source-backed — Every fact links to its origin document
  • MCP-native — Built-in Model Context Protocol server for AI assistants

Current Status

Stable:

  • Docker-based self-hosting
  • Local document import
  • Source-backed query API
  • Knowledge graph visualization
  • MCP server mode

Repository Layout

app/
  api/          FastAPI routes (sources, graph, query)
  cli/          CLI commands (ingest, query, graph, mcp)
  models.py     SQLAlchemy data model
  processing/   Extraction, embeddings
  services/     Ingest, query
  mcp/          MCP server implementation
frontend/       React 3-view app
tests/          Backend tests