agentic-retrieval

This skill should be used when indexing or searching a corpus with the pi-serini Lucene BM25 retriever, tuning BM25 k1/b parameters for long documents, handling missing Java 21 / pyserini dependencies, or when corpus scale demands a production-grade Lucene index.

Compare six retrieval strategies — contextual retrieval (TF-IDF/BM25/RRF fusion), its LLM-enriched lexical+ctx variant, turbovec (dense ANN), pi-serini (Lucene BM25), hybrid (lexical + dense fused with RRF), and tree-sitter (AST-boundary chunking with enclosing scope context) — with a stdlib-only offline core and graceful degradation when optional backends are missing. Use when the user wants to compare retrieval strategies on their project, search project files, or set up the retrieval engine.

This skill should be used when indexing or searching a code/script corpus with the tree-sitter retriever, wanting AST-boundary chunk spans plus an enclosing function/class breadcrumb on hits, or handling the RuntimeError raised when the treesitter extra is not installed.

This skill should be used when indexing or searching a corpus with the turbovec dense ANN retriever, choosing quantized embedding-based retrieval over lexical matching, handling the RuntimeError raised when the turbovec + local extras are not installed, or when keyword search misses paraphrases and synonyms (vocabulary-mismatch problems).

This skill should be used when combining lexical, dense, and Lucene retrieval rankings via reciprocal rank fusion, choosing which single retrieval method fits a task, deciding whether to add LLM/heuristic contextualization before indexing, or when a single retrieval method returns unsatisfying results and rankings should be combined.

This skill should be used when indexing or searching a corpus with the contextual lexical retriever (TF-IDF + BM25 fused with reciprocal-rank fusion), optionally enriching document text with LLM-generated context before indexing, or when a zero-dependency offline search over project files is needed.