Multi-Agent RAG

Live demo: multi-agent-rag.olegedly.com A config-driven multi-agent research system grounded in curated knowledge bases. Three specialist AI agents (Researcher → Critic → Synthesizer) collaborate via a LangGraph pipeline to answer questions — cited, grounded, and streamed live to a SolidJS dashboard. Powered by pgvector RAG, with a standalone MCP server for IDE integration.

Quick Start

First, create your config:

cp .env.example .env
# Then edit .env with your API keys

Then start the dev environment:

./dev.sh

This starts PostgreSQL (Docker), the FastAPI backend (hot-reload on localhost:8000), and the SolidJS frontend (Vite HMR on localhost:3000). The frontend proxies /api/* to the backend.

Prerequisites: Docker, uv (Python ≥3.13), Bun, pgweb (optional).

Architecture

SolidJS SPA → AG-UI/SSE → FastAPI → LangChain Agent Pipeline
                                        ├── Researcher Agent (corpus-scoped search)
                                        ├── Critic Agent (corpus-scoped verification)
                                        └── Synthesizer Agent (corpus-scoped answer)
                              LangChain RAG Tools (search_corpus, read_document)
                                   └── pgvector RAG (corpus-filtered)
                              MCP Server (standalone, for pi/Claude Desktop)
                                   └── pgvector RAG (corpus-filtered)

The system is corpus-scoped: each conversation binds to one corpus from start to finish. Every retrieval carries the active corpus_id. Adding a new knowledge base is an ingestion + config task.

Key modules

Module	Responsibility
backend/embeddings/	Abstract EmbeddingClient protocol with config-driven factory
backend/rag/	Chunker strategies + pgvector semantic search
backend/mcp_server/	Standalone MCP server: search_corpus, read_document
backend/agents/	LangGraph multi-agent pipeline: Researcher → Critic → Synthesizer
backend/main.py	FastAPI create_app() factory with DI for testing
frontend/src/	SolidJS SPA with conversation persistence

Configuration

Set via .env:

Variable	Purpose
LLM_PROVIDER	anthropic or openai
LLM_MODEL	Model name (e.g. deepseek-chat, gpt-4o)
LLM_API_KEY	API key
LLM_BASE_URL	API base URL
EMBEDDING_MODEL	Embedding model via OpenRouter
POSTGRES_*	PostgreSQL connection details

Adding a Corpus

1. Place documents in corpora/<slug>/
2. Add entry to backend/corpora.yaml (id, slug, name, chunker strategy, glob)
3. Run uv run scripts/seed_knowledge_base.py --corpus <slug>
4. Frontend picks it up from GET /api/corpora

Seeding

uv run scripts/seed_knowledge_base.py --all       # all corpora
uv run scripts/seed_knowledge_base.py --corpus eu-ai-act  # single corpus

Idempotent: SHA-256 diff against document_sources table (insert new, update changed, delete removed, skip unchanged).

Testing

make test           # full suite (backend + frontend)
make test_backend   # backend only
make test_frontend  # frontend only

All 422 tests pass (223 backend + 199 frontend) with 0 pyright/TypeScript errors.

Pure unit tests — no Docker or DB dependency for business logic. The create_app() DI makes mocking straightforward.

Deployment

Two Docker images, pushed to GHCR on main:

• ghcr.io/<owner>/multi-agent-rag-be:latest — FastAPI + LangChain + MCP
• ghcr.io/<owner>/multi-agent-rag-fe:latest — Caddy serving the SPA

CI/CD: GitHub Actions → build → push → Coolify webhook.