Search Settings

The [search] section controls search behavior defaults.

Configuration

[search]
default_limit = 10
semantic_weight = 0.7

Options

Semantic Weight

The semantic_weight parameter controls how hybrid search blends results:

• 1.0 = pure semantic search (vector similarity only)
• 0.0 = pure keyword search (full-text search only)
• 0.7 (default) = heavily favors semantic matches, with keyword results filling in exact-match gaps

The hybrid search uses Reciprocal Rank Fusion (RRF) to combine results from both search modes. See Architecture: Storage & Data Flow for details on the RRF algorithm.

Cross-encoder reranking (opt-in)

Absent by default. When configured, the top-K hybrid results are rescored by a local cross-encoder ONNX model after RRF fusion and before the final truncation:

[search.reranker]
model_path = "/models/cross-encoder.onnx"   # required, user-supplied
tokenizer_path = "/models/tokenizer.json"   # required, user-supplied
max_seq_len = 512                            # (query, passage) pair budget
top_k = 50                                   # results rescored per query
rerank_weight = 1.0                          # 1.0 = reranker replaces top-K order

• Bobbin never downloads reranker models. Point the paths at a cross-encoder exported to ONNX with a single relevance logit per (query, passage) pair. Missing paths refuse loudly at startup (bobbin serve) or at first hybrid search — never a silent downgrade.
• Blend rule: reranker logits are sigmoid-squashed to (0, 1); the K fused scores are min-max normalized; the final score is rerank_weight * sigmoid(logit) + (1 - rerank_weight) * fused_norm. At the default rerank_weight = 1.0 the reranker replaces the ordering within the top-K. Results beyond top_k keep their fused scores and positions.
• Applies to the hybrid mode of bobbin search, the HTTP /search endpoint, and the MCP search tool. Specialized lanes (beads, commits, archive search) are not reranked.
• Status note: the reranking stage is unit-tested with deterministic fake scorers; the ONNX model-in-the-loop path compiles and is seam-tested but has not been validated against a real cross-encoder model — treat retrieval quality as unmeasured until the eval harness runs it. Keep it off in production until then.