LCOD RAG Stack

This repository hosts the Retrieval-Augmented Generation (RAG) stack used by the LCOD organisation. It glues together the existing AI services (Ollama, Open WebUI and Dify) with a dedicated ingestion pipeline and a lightweight retrieval API.

Goals

• Centralise knowledge from LCOD repositories (lcod-spec, lcod-resolver, lcod-components, lcod-kernel-*, …) in a searchable vector store.
• Expose a simple API that other tools (Open WebUI, Dify apps, IDE assistants) can call for semantic search + answer generation.
• Automate ingestion so new commits or documentation drops can be synchronised quickly.
• Remain close to the existing infrastructure (Docker, Ollama, Weaviate/Qdrant) to simplify operations.

The stack is meant to run on the shared LCOD infrastructure, but every component is containerised so the same setup can be reproduced elsewhere.

Repository layout

app/
  rag_api/        # FastAPI service (retrieval + generation)
config/
  .env.example          # Example environment file for the API container
docs/
  ARCHITECTURE.md # High-level description of the stack and integration points
  OPERATIONS.md   # Operational playbooks (deployment, ingestion, backups)
Dockerfile        # Builds the rag-api container image
docker-compose.yml
Makefile          # Automation shortcuts (install, run, ingest, lint)
requirements.txt  # Python dependencies for the API service
packages/rag/components/ingest.*  # LCOD ingestion pipeline

Quick start (local workstation)

1. Create and activate a virtualenv, then install dependencies:

   python -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt

2. Copy the sample environment file and edit it to match your setup:

   cp config/.env.example config/.env

3. Launch the services with Docker Compose:

   docker compose up -d

4. Run an ingestion pass with lcod-run (download the latest LCOD runtime bundle first):

   lcod-run --compose packages/rag/components/ingest.run_pipeline/compose.yaml

   The pipeline honours the same environment variables as the previous Python scripts: set OLLAMA_BASE_URL, QDRANT_URL, QDRANT_COLLECTION, QDRANT_API_KEY and RAG_EMBED_MODEL as needed before running the command.

5. Query the API:

   curl -X POST http://localhost:8088/query \
        -H 'Content-Type: application/json' \
        -d '{"query": "What is the resolver responsible for?"}'

Registry snapshot powered by lcod-run

The RAG ingestion pipeline ships LCOD components that can be executed directly with the standalone lcod-run binary (Rust). Once the repository is cloned, the registry snapshot can be refreshed without extra configuration:

lcod-run --compose packages/rag/components/registry.refresh_snapshot/compose.yaml

The compose downloads the official LCOD catalogues, resolves the components, and updates data/registry.snapshot.json only when the content changes. When a different catalogue pointer or snapshot location is required, provide overrides through the compose state (see component READMEs for the available fields). Standard lcod-run flags (--log-level, --timeout, --global-cache, …) remain available for advanced scenarios.

Deployment

The target server already runs the required dependencies:

• Ollama (HTTP API on http://127.0.0.1:11434) for both embeddings and generation.
• Traefik + Docker for routing and SSL.
• Existing AI stacks (Dify, Open WebUI) that will consume this RAG API.

See docs/OPERATIONS.md for the detailed procedure (clone repo, configure compose, schedule ingestion, monitoring).

Status

The repository currently provides the base scaffolding, ingestion utilities and API skeleton. Next iterations will plug the API into Open WebUI (custom retriever) and Dify datasets.