rag-llm-infra 0.1.1

Vendor-neutral RAG + LLM serving infrastructure: swappable LLM protocol and vector store (FAISS/NumPy/Qdrant), cached embedding index, and observability.

RAG + LLM Serving Infrastructure

An installable, vendor-neutral foundation for retrieval-augmented LLM applications: a swappable vector store, a cached embedding index, a provider-agnostic LLM protocol, the observability around them, a FastAPI serving layer, and a retrieval-quality eval gate.

Distilled infrastructure layer — typed, tested, packaged, and runnable on its own.

Install

pip install rag-llm-infra                                   # core (numpy)
pip install "rag-llm-infra[faiss,qdrant,openai,serve]"      # + native backends, OpenAI, serving
pip install -e ".[dev]"                                     # from a local clone, for development

Quickstart — end-to-end RAG (no API key, no network)

git clone https://github.com/MarwaBS/rag-llm-infra && cd rag-llm-infra
pip install -e .
python example.py

embed documents → index in a VectorStore → retrieve top-k for a query
                → build a grounded prompt → answer with an LLMProtocol backend

Runs on the NumPy vector store + the deterministic mock LLM, so it needs no key. In production, swap the demo embedder for EmbeddingEngine and get_llm("mock") for get_llm("openai").

Serve it

pip install "rag-llm-infra[serve]"
uvicorn rag_llm_infra.serve:app          # or: docker build -t rag-llm-infra . && docker run -p 8000:8000 rag-llm-infra

curl -XPOST localhost:8000/index -d '{"documents":["FAISS is in-process vector search","Qdrant is a vector database"]}' -H 'content-type: application/json'
curl -XPOST localhost:8000/query -d '{"query":"vector search","k":1}'      -H 'content-type: application/json'

What's inside

Module	Responsibility
rag_llm_infra.llm_protocol	LLMProtocol — runtime_checkable Protocol over OpenAI / Anthropic-stub / Mock; factory get_llm()
rag_llm_infra.vector_store	VectorStoreProtocol — in-process FAISS IndexFlatIP, pure-NumPy fallback, real Qdrant (batched search)
rag_llm_infra.evidence_index	EmbeddingEngine — SentenceTransformers embeddings + a memory-pressure-aware cache (insertion-order eviction) guarded by a writer-preferring reader/writer lock, so the slow model.encode runs outside the lock
rag_llm_infra.tracing	OpenTelemetry spans with console-exporter + no-op fallbacks
rag_llm_infra.log_config	structured JSON logging + an llm_call latency/token timer
rag_llm_infra.serve	FastAPI service (/index, /query, /health) wiring the parts together
rag_llm_infra.faithfulness	groundedness(answer, contexts) — lexical faithfulness metric for RAG output
rag_llm_infra.fallback	FallbackLLM — budget-aware multi-provider routing; drop-in LLMProtocol

Quality gates

python -m eval.retrieval_eval      # recall@1 / MRR — retrieval mechanics over the demo embedder
python -m eval.generation_eval     # groundedness (faithfulness) of generated answers

Both run in CI: a retrieval regression (recall@1 ≥ 0.80, MRR ≥ 0.85) or a faithfulness regression (grounded answer below threshold, or the metric failing to flag a hallucinated control) fails the build and cannot merge.

Engineering principles demonstrated

• Swap by interface — LLMProtocol / VectorStoreProtocol make the model and the index runtime-swappable.
• Degrade, don't crash — FAISS / Qdrant / OpenTelemetry / SentenceTransformers are lazily imported with working fallbacks; missing infra never hard-fails import.
• Measured, not asserted — a retrieval eval gate, not just unit tests; packaged and CI-built end to end.

Develop / test

pip install -e ".[dev]"     # installs FAISS + Qdrant + serve extras too
ruff check . && pytest && python -m eval.retrieval_eval

CI installs the native backends, so the FAISS and Qdrant tests run there (they skip only when those libraries are absent).

License

MIT — see LICENSE.