From bedebc3a2a5b28baa1c81f8017e8a50e3faf9b55 Mon Sep 17 00:00:00 2001 From: Wagner Lacerda Date: Tue, 20 Oct 2026 17:23:43 -0300 Subject: [PATCH 1/2] docs(readme): update the RAG entry for semantic and hybrid retrieval --- README.md | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index bc21e78..04718cb 100644 --- a/README.md +++ b/README.md @@ -39,9 +39,10 @@ side by side through a single interface. End-to-end retrieval-augmented generation pipeline that answers questions grounded in a document set, with citations back to the source chunks. -**Stack:** `TF-IDF retrieval` `FastAPI` `Claude` +**Stack:** `Hybrid retrieval (RRF)` `Qdrant` `fastembed` `TF-IDF` `FastAPI` `Claude` -- TF-IDF retrieval chosen deliberately over embeddings — runs fully offline, deterministic, no API key required just to test retrieval quality +- Retrieval is chosen by configuration — TF-IDF, semantic (embeddings in Qdrant), or hybrid, which merges both rankings with Reciprocal Rank Fusion; TF-IDF stays a first-class mode that runs fully offline with no model download +- Each mode is measured on a 40-question golden set (hit rate@3: hybrid 0.975, semantic 0.900, TF-IDF 0.875), and CI fails if hybrid falls below the TF-IDF baseline — the README also says where semantic search alone loses - Every generated answer returns alongside the exact chunks it was grounded in, not just a confident-sounding response --- From a9231e295c91c19ed6f2564b51690929913feeba Mon Sep 17 00:00:00 2001 From: Wagner Lacerda Date: Tue, 20 Oct 2026 17:23:48 -0300 Subject: [PATCH 2/2] docs(roadmap): update the RAG row for semantic and hybrid retrieval --- ROADMAP.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ROADMAP.md b/ROADMAP.md index 14e2e53..411bdf5 100644 --- a/ROADMAP.md +++ b/ROADMAP.md @@ -25,7 +25,7 @@ | Project | Status | Stack | What it demonstrates | |---|---|---|---| | [Multi-Provider MCP Server](https://github.com/LacerdaTraderCode/mcp-multi-llm-server) | ✅ | MCP (official SDK) · Claude/GPT/Gemini adapters · httpx2 | MCP in practice + provider-agnostic architecture | -| [RAG Knowledge Assistant](https://github.com/LacerdaTraderCode/rag-knowledge-assistant) | ✅ | TF-IDF retrieval · FastAPI · Claude | End-to-end RAG with cited, grounded answers | +| [RAG Knowledge Assistant](https://github.com/LacerdaTraderCode/rag-knowledge-assistant) | ✅ | Hybrid retrieval (TF-IDF + embeddings, RRF) · Qdrant · fastembed · FastAPI · Claude | End-to-end RAG with cited, grounded answers, and retrieval measured per mode on a golden set with a CI quality gate | | [AI Data & Evaluation](https://github.com/LacerdaTraderCode/ai-data-evaluation) | ✅ | Anchored rubrics · LLM-as-judge · preference data | Rubric design, golden-set agreement checking, DPO-shaped preference data | | [Price Prediction with ML](https://github.com/LacerdaTraderCode/price-prediction-ml) | ✅ | scikit-learn (RandomForest + MLP) | Honest ML methodology: lookahead-free features, chronological split, baseline comparison |