freecomai-hybrid-rag

Hybrid retrieval-augmented generation — graph × vectorstore → enriched context.

Status

🚧 Placeholder release — v0.0.0. Full API lands in v0.1.0 (planned Q2-Q3 2026).

This package reserves the name freecomai-hybrid-rag on PyPI. Do not install for production yet.

What it will do

A domain-agnostic Python library that takes two inputs — a knowledge graph (NetworkX) and a vectorstore (ChromaDB / Qdrant / Pinecone / ...) — and returns enriched search results that combine:

• Dense retrieval (semantic similarity top-K chunks)
• Graph context (entities mentioned in those chunks + their 1-hop neighbors)
• Cluster summary (dominant community the result falls into, with LLM-generated description)
• Related bugs / decisions / events from the same cluster

One function call → not just "here are 5 similar chunks", but "here is the full contextual map around your query".

Why

Vector RAG alone misses structural relationships. GraphRAG (Microsoft pattern, 2024) adds Leiden clustering + community reports but loses fast semantic retrieval. HybridRAG (arXiv 2408.04948, NVIDIA + BlackRock, 2024) fuses both — proven superior on financial QA tasks.

This library makes that pattern plug-and-play for any domain: your graph, your vectorstore, our enrichment logic.

Planned API (v0.1.0)

from hybrid_rag import enriched_search, EnrichedResult

result: EnrichedResult = enriched_search(
    graph=my_networkx_graph,
    vectorstore=my_chroma_client,
    query="how did we decide on architecture X?",
    k=5,
)

print(result.dominant_cluster.summary)
for chunk in result.chunks:
    print(chunk.text, chunk.score)
for bug in result.related_bugs:
    print(bug.name, bug.status)

Roadmap

• v0.0.0 (placeholder) — name reservation
• v0.1.0 (Q2-Q3 2026) — core enriched_search() function
• v0.2.0 — pluggable cluster summarizers (Claude / GPT / local LLMs)
• v0.3.0 — optional reranker integration (bge-reranker-v2-m3, Cohere, Jina)
• v1.0.0 — stable API + comprehensive docs

Author

Built by FreeComAI, Astana, Kazakhstan. Founder: Bolatbek Barmagambetov (barmagambetov.b@gmail.com).

License

Apache 2.0 — see LICENSE.