semantic-search

name	semantic-search
description	Configure, build, and verify dagayn embeddings and hybrid semantic search without losing FTS fallback behavior.
argument-hint	[query]

Semantic Search

Use this when semantic search quality, embedding setup, or hybrid search status matters.

Installed Search Mode

This packaged skill is mode-neutral. dagayn install rewrites this section with the selected embedding mode so agents can avoid stale or wasteful search advice.

Workflow

1. Start with graph freshness:

   dagayn tool list_graph_stats_tool
   

2. Run the intended search and read search_mode / per-result source:

   dagayn tool semantic_search_nodes_tool --arg query='"auth handler"' --arg detail_level='"minimal"'
   

   search_mode="hybrid" means embeddings and FTS were merged. fts_only is still valid, but semantic recall is lower. embedding_only means the vector path ran while FTS was unavailable, and keyword_fallback means the FTS index was absent and only LIKE matching ran.
   • For exact symbol/name lookup, trust exact name / qualified_name hits even when search_mode is fts_only.
   • For fuzzy purpose searches, prefer high-ranked hits with source="both" or source="embedding".
   • For process-pattern prose, check rerank_intent="process_pattern"; this routes to narrative embeddings when they are available.
   • Use per-result source to explain why a hit ranked: fts, embedding, both, or keyword.
3. Hand off the best hit to graph tools:
   • If exactly one result matches the intended name or qualified_name, use that qualified_name for query_graph_tool.
   • If several fuzzy hits look plausible, inspect the top 3 with query_graph_tool(pattern="file_summary") or a relationship query before drawing conclusions.
   • If the user wants callers, callees, tests, docs, imports, or children, use query_graph_tool; use traverse_graph_tool only for a bounded neighborhood after the start node is clear.
   • If the search result has a next_action, follow it before widening the query.
4. If embeddings are missing or stale, build them through the graph tools:
   • Incremental local refresh: build_or_update_graph_tool(local_embedding="bge-m3")
   • Dedicated embedding pass: embed_graph_tool
   • Full local refresh: build_or_update_graph_tool(full_rebuild=True, local_embedding="bge-m3") only when explicitly doing embedding-quality or end-to-end maintenance work. Before any embedding-enabled full rebuild, state the reason and get explicit confirmation from the user; do not use it for parser, flow, documentation, or ordinary implementation verification.
5. For CLI fallback:

   dagayn build --local-embedding
   dagayn update --local-embedding
   dagayn tool embed_graph_tool
   

6. Re-run the same semantic_search_nodes_tool query and compare result count, search_mode, rerank_intent, and whether high-value hits now have source="embedding" or source="both".

Troubleshooting

• fts_only is acceptable for exact symbol/name lookup; do not rebuild embeddings just to find a precise identifier.
• keyword_fallback means the FTS index is absent; refresh graph/FTS before drawing conclusions about search quality.
• provider_mismatch or missing_vectors in embedding health means the query did not find matching vectors for the selected provider/text mode. Refresh embeddings only when the task actually needs fuzzy recall.
• Use local BGE-M3 for reusable developer environments when embeddings are useful; use FTS-only when startup time or memory is tight.
• If Qwen sidecar startup fails, check the local server binary (auto or llama-server), port, and timeout before changing graph data.
• If provider imports are unavailable, keep going with FTS and report the reduced recall instead of blocking unrelated work.

Efficiency Rules

• Use FTS-only results for exact names; use material embeddings for fuzzy purpose concepts; use narrative embeddings/process-pattern intent for queries about behavior such as calls, reads, writes, loops, returns, or merges.
• Treat semantic search as start-node discovery, not final proof. Confirm behavior, relationships, or coverage with query_graph_tool, flow_tool, review_tool, or source reads.
• Do one before/after query to prove search quality changed. Do not rebuild the graph repeatedly without a changed file set or a failed verification.
• Never use an embedding-enabled full rebuild to compensate for untracked files not appearing in graph queries. Stage or otherwise expose the files first, then run the smallest non-embedding graph refresh that proves the claim.