docscontext

DocsContext

A pure Go (CGO_ENABLED=0) GraphRAG tool inspired by Microsoft GraphRAG. Ingests unstructured documents, builds a hierarchical knowledge graph with community detection, and exposes an MCP server + embedded Web UI on a single port.

Features

• GraphRAG pipeline — 5-phase: load → chunk → embed → graph extraction → community detection
• Knowledge graph — entity/relationship/claim extraction via LLM (JSON mode)
• Louvain community detection — pure Go, hierarchical, no external dependencies
• Three LLM providers — Azure OpenAI, Ollama (local), HuggingFace TGI
• 12 MCP tools — local search, global search, graph walk, community reports, and more
• Embedded Web UI — vis-network graph explorer, semantic search, document browser
• Single binary — zero CGO, cross-compiles to Linux / macOS / Windows

Install

go install github.com/RandomCodeSpace/docscontext@latest

Or build from source:

git clone https://github.com/RandomCodeSpace/docscontext.git
cd DocsContext
CGO_ENABLED=0 go build -o docscontext .

Quick Start

# 1. Create config
mkdir -p ~/.docscontext
cp config.example.yaml ~/.docscontext/config.yaml
# Edit ~/.docscontext/config.yaml — set your LLM provider

# 2. Index documents (Phases 1-2: load, chunk, embed, extract entities)
docscontext index ./your-docs/ --workers 4

# 3. Build knowledge graph (Phases 3-4: community detection + summaries)
docscontext index --finalize

# 4. Check stats
docscontext stats

# 5. Start server
docscontext serve --port 8080

Open http://localhost:8080 for the Web UI.

Configuration

Copy config.example.yaml to ~/.docscontext/config.yaml and edit:

data_dir: ~/.docscontext/data

llm:
  provider: ollama          # azure | ollama | huggingface

  ollama:
    base_url: http://localhost:11434
    chat_model: llama3.2
    embed_model: nomic-embed-text

  azure:
    endpoint: https://myresource.openai.azure.com
    api_key: ${AZURE_OPENAI_API_KEY}
    api_version: "2024-02-01"
    chat_model: gpt-4o
    embed_model: text-embedding-3-small

indexing:
  chunk_size: 512
  chunk_overlap: 50
  workers: 4

server:
  host: 127.0.0.1
  port: 8080

Environment variable overrides use the docscontext_ prefix:

docscontext_LLM_PROVIDER=azure
docscontext_LLM_AZURE_API_KEY=sk-...
docscontext_SERVER_PORT=9090

CLI

# Index a file or directory
docscontext index ./docs/ [--force] [--workers 4] [--verbose]

# Run community detection + LLM summaries
docscontext index --finalize

# Show statistics
docscontext stats
docscontext stats --json

# Start MCP + Web UI server
docscontext serve [--port 8080] [--host 127.0.0.1]

MCP Tools

Connect any MCP client to http://localhost:8080/mcp/sse.

REST API

GET  /api/stats
GET  /api/documents
GET  /api/documents/{id}
POST /api/search          {"query":"...","mode":"local|global","top_k":5}
GET  /api/graph/neighborhood?entity=<name>&depth=2
GET  /api/entities
GET  /api/communities
GET  /api/communities/{id}
POST /api/upload

Architecture

Document In
    │
    ▼ Phase 1 — Text Units
  Loader (PDF/DOCX/TXT/MD) → Chunker → Embedder → SQLite

    ▼ Phase 2 — Graph Extraction  [parallel per document]
  LLM → Entities + Relationships + Claims → SQLite

    ▼ Phase 3 — Community Detection  [post-index finalization]
  Louvain algorithm → hierarchical community assignments

    ▼ Phase 4 — Community Summaries  [parallel]
  LLM → CommunityReport → embed summary → SQLite

    ▼ Phase 5 — Structured Doc
  LLM → JSON summary → SQLite

All data lives in a single SQLite file at $DATA_DIR/docscontext.db.

Supported File Types

Tip: For best graph quality, prefer documents with clear structure (headings, named entities, factual prose). Scanned PDFs or heavily formatted spreadsheets will produce sparse graphs.

License

MIT