CRG hybrid-search provider mismatch bug. code_review_graph/search.py:184 instantiates EmbeddingStore(store.db_path, model=model) for query-time encoding, ignoring the provider column already stored in the embeddings table per row. Result: when nodes were embedded with cloud (Google text-embedding-004, 768-dim), query encoding still tries LocalEmbeddingProvider (sentence-transformers, 384-dim) and fails with "sentence-transformers not installed". Bench on 2026-05-04 confirmed: 1340 Google embeddings stored in .code-review-graph/graph.db, but every hybrid query logs Embedding search failed: sentence-transformers not installed and falls back to FTS5-only. Two fixes possible: (a) patch _embedding_search in search.py to read the stored provider from the embeddings table before instantiating the query-side EmbeddingStore (~10 lines, PR upstream to https://github.com/tirth8205/code-review-graph), or (b) install pip install code-review-graph[embeddings] and re-embed with local all-MiniLM-L6-v2 (sentence-transformers + pytorch ~700 MB into pyenv 3.13.12). Until fixed, vector search is dormant and semantic_search_nodes_tool is FTS5-only on this repo — which is empirically still good for symbol queries (see AGENTS.md "Practical query routing"), so this is not blocking, just suboptimal.
CRG hybrid-search provider mismatch bug.
code_review_graph/search.py:184instantiatesEmbeddingStore(store.db_path, model=model)for query-time encoding, ignoring theprovidercolumn already stored in theembeddingstable per row. Result: when nodes were embedded with cloud (Googletext-embedding-004, 768-dim), query encoding still triesLocalEmbeddingProvider(sentence-transformers, 384-dim) and fails with "sentence-transformers not installed". Bench on 2026-05-04 confirmed: 1340 Google embeddings stored in.code-review-graph/graph.db, but every hybrid query logsEmbedding search failed: sentence-transformers not installedand falls back to FTS5-only. Two fixes possible: (a) patch_embedding_searchinsearch.pyto read the stored provider from the embeddings table before instantiating the query-sideEmbeddingStore(~10 lines, PR upstream to https://github.com/tirth8205/code-review-graph), or (b) installpip install code-review-graph[embeddings]and re-embed with localall-MiniLM-L6-v2(sentence-transformers + pytorch ~700 MB into pyenv 3.13.12). Until fixed, vector search is dormant andsemantic_search_nodes_toolis FTS5-only on this repo — which is empirically still good for symbol queries (seeAGENTS.md"Practical query routing"), so this is not blocking, just suboptimal.