OmniSearch

Internal RAG (Retrieval-Augmented Generation) system for organizational knowledge bases.

📄 Architecture decisions · 🗺️ Roadmap

Stack

Component	Technology
Vector DB	Qdrant (self-hosted, Docker)
Embeddings	BAAI/bge-large-en-v1.5 (HuggingFace)
Orchestration	LangChain + LCEL
LLM	OpenAI (default: gpt-4o-mini) · Anthropic Claude (claude-3-5-haiku) — swappable via LLM_PROVIDER env var
Backend	FastAPI + SSE streaming
Frontend	Chainlit
CI/CD	GitHub Actions

Quick Start

1. Configure environment

cp .env.example .env
# Edit .env with your API keys

2. Start the full stack

docker compose up --build
# Frontend available at http://localhost:8501

3. Run the ingestion pipeline

# Drop PDFs into data/sources/ then:
docker compose --profile ingestion run ingestion

# Or trigger manually via GitHub Actions → workflow_dispatch

Development

# Install all dependencies
pip install -e ".[ingestion,backend,frontend,dev]"

# Start services with hot reload
docker compose -f docker-compose.yml -f docker-compose.dev.yml up

# Run tests
pytest tests/unit/ -v
pytest tests/integration/ -v   # Requires Qdrant running

Architecture

flowchart TD
    GHA["GitHub Actions\n(indexer.yml)"] --> Pipeline["ingestion/pipeline.py"]
    Pipeline --> Conn["connectors/\nPDF · Notion · Confluence"]
    Pipeline --> Chunker["chunker.py"]
    Pipeline --> Embedder["embedder.py"]
    Pipeline --> VS["vector_store.py"]
    VS --> Qdrant[("Qdrant collection\nomnisearch")]
    Qdrant --> Main["backend/main.py\nFastAPI"]
    Main --> Query["/api/v1/query"]
    Query --> Ret["retriever.py"]
    Query --> Chain["chain.py"]
    Query --> SSE["SSE streaming"]
    SSE --> Frontend["frontend/app.py\nChainlit"]

Loading

Data Sources

Source	Connector	Env Vars Required
Local PDFs	PdfConnector	PDF_SOURCE_DIR
Notion	NotionConnector	NOTION_TOKEN, NOTION_ROOT_PAGE_ID
Confluence	ConfluenceConnector	CONFLUENCE_URL, CONFLUENCE_USER_EMAIL, CONFLUENCE_TOKEN, CONFLUENCE_SPACE_KEY

Ingestion Automation

The indexer runs automatically:

• Nightly at 2AM UTC (cron schedule)
• On push when files are added to data/sources/
• Manually via GitHub Actions → Run workflow (choose source type)

Security

• Only the frontend port (8501) is exposed to the host
• All API keys are stored as GitHub Secrets / Docker env vars
• Qdrant and the backend communicate over an internal Docker network
• The LLM is instructed to cite sources and refuse to answer outside the knowledge base

Stress Testing

# Place complex PDFs in tests/stress/fixtures/ then:
RUN_STRESS_TESTS=1 pytest tests/stress/ -v

Cost & Latency by Provider

The LLM backend is swappable via LLM_PROVIDER (see ADR 0001). The table below estimates the cost of a typical RAG query (~1.5k tokens of context + ~300 tokens of output), using public list pricing (per 1M tokens, as of Jun 2026). Latency varies with network and load.

Provider	Model	Input $/1M	Output $/1M	~Cost/query	~Latency (p50)
OpenAI	gpt-4o-mini	$0.15	$0.60	~$0.0004	~1–2s
Anthropic	claude-3-5-haiku	$0.80	$4.00	~$0.0024	~1–2s

List prices, not negotiated rates. For real measurements in your environment: python scripts/bench_providers.py (needs API keys; runs a few prompts per provider and reports measured latency, token usage, and estimated cost).

ADR — Why This Stack

Formal architecture decisions live in docs/adr/. Stack highlights:

BGE embeddings over OpenAI embeddings — self-hosted, no per-token cost, strong multilingual performance for organizational knowledge bases.

Qdrant over pgvector — purpose-built vector DB with filtering, payload indexing, and horizontal scaling; pgvector is adequate for small workloads but operationally simpler to replace than retrofit.

Chainlit over a custom frontend — ships streaming, source citation UI, and auth out of the box; building equivalent features from scratch would cost 2–3 weeks with no differentiated value.

Provider abstraction (OpenAI ↔ Anthropic) — a thin LLMProvider interface keeps the model backend a config choice, not a code dependency. See ADR 0001.