Fuse BM25 and GIST seeds with RRF and ColBERT reranking for GraphRAG retrieval.
npx skills add https://github.com/thistleknot/skills --skill gist-retriever
Please help me install this Agent Skill. Skill: gist-retriever Source: https://github.com/thistleknot/skills/tree/main/gist-retriever Command: npx skills add https://github.com/thistleknot/skills --skill gist-retriever
This Skill solves the problem of reliably retrieving relevant evidence from a large knowledge corpus by combining lexical, dense, and graph-adjacent neighborhood expansion into a single candidate pipeline.
Use gist-retriever to retrieve a high-quality evidence candidate set for a query before running syllogistic reasoning or answer synthesis.
High-intent search queries and answers about installing and using this skill.
Hybrid BM25 and dense retrieval combines lexical and semantic results using Reciprocal Rank Fusion (RRF). This preserves ranking diversity while building a single candidate pool for downstream neighborhood expansion and reranking.
GraphRAG neighborhood expansion takes fused BM25 and dense seeds and maps them into a local L2 semantic space using triplet expansion and dense centroids. It is needed when flat vector search misses conceptually adjacent context.
ColBERT late interaction reranking improves RAG precision by applying token-level scoring to expanded neighborhood candidates. It refines the fused pool after L2 expansion, ensuring higher-precision final ranking before candidate reconstruction.
Yes, gist-retriever executes locally over a compact semantic triplet space using Chroma-based vector search. It routes queries through a tiered cascade from wiki and markdown to local memory, making it suitable for agentic coding.
The best way to fuse lexical and dense search results is Reciprocal Rank Fusion (RRF). It merges BM25 lexical seeds with dense GIST semantic seeds into a unified candidate pool, preserving ranking diversity before L2 neighborhood expansion.
Relying solely on BM25 for corpus retrieval misses semantic context and fails to capture conceptually adjacent evidence. Expanding fused seeds into an L2 neighborhood with dense centroids and ColBERT reranking overcomes these lexical matching limitations.
drshailesh88Index local documents and search them with BM25 and vector embeddings.
hightempEnforce evidence-based reasoning and explicit uncertainty in AI-generated answers.
ArogyaReddyCombine vector similarity and keyword matching with RRF, Linear, Cross-encoder, or Cascade fusion.
zechenzhangAGIPerform billion-scale k-NN similarity search with FAISS index types and GPU acceleration.
