v0.1.0-alpha.0 - Green Infinuum

RAG Loom - RAG Platform Kit v0.1.0-alpha.0 - Green Infinuum – Release Notes

• Overview

  • FastAPI microservice implementing Retrieval-Augmented Generation (RAG): document ingestion → chunking → embedding → vector store → semantic search → LLM generation with retrieved context.

• Key Features

  • Ingestion: Upload PDF/TXT, clean, chunk (sliding window), batch-embed, and store in vector DB.
  • Semantic Search: Vector similarity search with configurable top_k and similarity threshold (with fallback).
  • RAG Generation: Retrieves top-K relevant chunks and calls an LLM with grounded context; returns answer + sources.
  • Pluggable Providers:
    • Embeddings: Sentence-Transformers (local, default), OpenAI (text-embedding-), Cohere (embed-english-).
    • Vector Stores: ChromaDB (local), Qdrant (HTTP/gRPC), Redis (vector index).
    • LLMs: Ollama (default), OpenAI, Cohere, Hugging Face Transformers.
  • Black-box E2E Tests: Real HTTP tests against infra (Qdrant, Ollama) started via docker compose.
  • Unit and Integration tests for functions used in endpoints
  • CI/CD: GitHub Actions workflow to run the e2e suite; caches Docker images to speed up runs.

• API Endpoints

  • GET /health: Service and platform metadata (vector store, embedding model, llm provider).
  • POST /api/v1/ingest: Multipart upload of PDF/TXT → returns chunks_created and document_id.
  • POST /api/v1/search: Body: { query, top_k, similarity_threshold, filters? } → list of relevant chunks.
  • POST /api/v1/generate: Body: { query, search_params?, context?, temperature, max_tokens } → answer + sources.

• Configuration

  • Env-driven via app/core/config.py and .env:
    • VECTOR_STORE_TYPE: chroma | qdrant | redis
    • CHROMA_PERSIST_DIRECTORY
    • QDRANT_URL, QDRANT_API_KEY
    • REDIS_URL
    • EMBEDDING_MODEL, EMBEDDING_DIM (default 384)
    • LLM_PROVIDER: ollama | openai | cohere | huggingface
    • OLLAMA_BASE_URL, OLLAMA_MODEL
    • TOP_K, SIMILARITY_THRESHOLD
    • SERVICE_HOST, SERVICE_PORT
  • Defaults: Qdrant at http://localhost:6333, Ollama at http://localhost:11434, local sentence-transformers.

• Requirements

  • Python 3.12
  • Docker + Docker Compose (for infra/e2e)
  • For OpenAI/Cohere/HF providers, set the relevant API keys in .env.

• Quick Start (Service)

  • Local venv (3.12) + run:
    • python3.12 -m venv venvpy312 && source venvpy312/bin/activate
    • pip install -r requirements.txt
    • uvicorn app.main:app --host 0.0.0.0 --port 8000
  • Or use helper:
    • bash utilscripts/quick_start.sh start
    • bash utilscripts/quick_start.sh stop

• Quick Start (Infra + E2E)

  • Start infra + run tests (black-box):
    • bash utilscripts/e2e_run.sh
  • What it does:
    • docker compose up qdrant, redis, ollama (under tests/)
    • waits on health (Qdrant, Ollama), ensures a small Ollama model is pulled
    • sets up Python 3.12 venv, installs tests/requirements-test.txt
    • starts the API locally, waits for /health
    • runs pytest -m e2e
    • stops API and stops containers (keeps them cached for faster re-runs)

• CI/CD

  • .github/workflows/e2e.yml:
    • Caches Docker images (Qdrant, Redis, Ollama) between runs.
    • Executes utilscripts/e2e_run.sh.
    • On failure, uploads docker compose logs artifact.
    • Tears down infra when finished.
  • .github/workflows/test.yml:
  • for unit and integration testing

• Notes and Limitations

  • Default small Ollama model used for CI speed; switch OLLAMA_MODEL for higher quality.
  • For OpenAI/Cohere/HF, provide API keys; otherwise the local embedding/LLM options are used.
  • Retrieval applies threshold; if no result passes, it falls back to top results to avoid empty context.

Project Summary
Documentation
Tests