Skip to content

v1.4.0

Choose a tag to compare

View all tags
@prih prih released this 23 Mar 13:47
· 175 commits to main since this release
v1.4.0
2151461

Released: March 2026

Highlights

  • Code-Optimized Embedding Model — code graph now defaults to jinaai/jina-embeddings-v2-base-code via new codeModel config field. Separate inheritance chain: graphs.code.model → project.codeModel → server.codeModel → code defaults.
  • Full Body in Code Embeddings — code symbols now embed signature + docComment + body (was signature + docComment only). Functions without JSDoc are now visible to semantic search.
  • Edge-Specific BFS Decay — code graph BFS uses per-edge-type decay: contains (0.95), extends/implements (0.85), imports (0.70). Reflects that class→method is a tighter relationship than a cross-file import.
  • Hybrid File Search — file-level searches (search_files, search_topic_files, search_all_files) now use BM25 + vector hybrid (was vector-only). Exact filename queries like "embedder.ts" now work reliably.
  • Embedding API Model Selection — POST /api/embed accepts model: "default" | "code" to select which embedding model to use. Both models loaded at startup when embeddingApi is enabled.
  • Graph Data Versioning — persisted graphs now store GRAPH_DATA_VERSION. Version mismatch triggers automatic re-index (alongside existing embedding fingerprint check).

Search Improvements

  • BFS queue.shift() replaced with index pointer — O(1) dequeue instead of O(n) array shift
  • File paths normalized for embedding: src/lib/search/code.ts → src lib search code ts for better tokenization
  • embedding.maxChars default raised from 8000 to 24000, matching ~8k token model capacity

Configuration

  • New codeModel field at server/project/workspace levels with its own inheritance chain
  • New embedding.remoteModel field: "default" or "code" — auto-set to "code" for code graph with remote embedding
  • New CODE_EDGE_DECAY constants in defaults for per-edge-type BFS decay
  • GRAPH_DATA_VERSION = 2 — bump when changing embedding content or stored format

Breaking Changes

  • Code graph default model changed from Xenova/bge-m3 to jinaai/jina-embeddings-v2-base-code — existing code graphs will be automatically re-indexed on first startup
  • embedding.maxChars default changed from 8000 to 24000
  • Embedding API embeddingApiModelName option replaced with embeddingApiModelNames: { default, code }
Assets 2
Loading