First met in aicp-c3-session-22 · Dev libraries & frameworks
Open-source AI orchestration framework for building production-grade RAG pipelines and agents — the LEGO blocks under a custom answer engine.
You want a self-hostable, provider-swappable RAG core (OpenAI for demo, Ollama for production) rather than a hosted RAG service.
Tiny real footprint: ~107MB installed with embedded Qdrant in Session 22's measured comparison — while RAGflow demanded 16GB RAM + 50GB disk
Provider-agnostic by design — the same pipeline ran OpenAI and Ollama adapters
Pairs with an embedded vector store, so single-tenant RAG needs no database service at all
Orchestrates the full RAG cycle: ingestion, chunking, embedding, retrieval, grounded generation
Ships modular pipeline components you compose instead of writing retrieval plumbing
Backs the Session 22 evidence gate: <=6 retrieved chunks, citations before generation, refuse-on-empty
Doesn't include a UI — Talent Lens needed its own Next.js front end and FastAPI service around it
Doesn't pick your model or vector store — those are adapter decisions you still own
Isn't a hosted service — you run it, which is the point
The orchestration layer of a self-hosted enterprise RAG: Haystack + embedded Qdrant under FastAPI, any LLM behind an adapter.
Dev libraries & frameworks
Strong candidate for any future 'ask my knowledge base' feature over the course KB; no current build uses it.
The Session 22 evidence-gate spec (citations-first, refuse-on-empty, untrusted-document labeling) is a ready acceptance bar if the KB site ever grows a Q&A feature.
current
qdrant: Session 22's chosen pairing — embedded Qdrant as the vector store
ollama: The production-grade local-LLM provider in the enterprise build
LangChain / LangGraph: Overlapping orchestration territory; Haystack was the course's RAG-specific pick
| Session | Coverage | Moment | What happened |
|---|---|---|---|
| Session 22: Automate Admin Tasks — Advanced RAG (Enterprise 'Internal Perplexity' Built Live) | demonstrated | Chosen RAG core — ~107MB with embedded Qdrant; he also wants it inside his Hermes agent |