Skip to main content
Search is an offline index over seller-declared Bazaar metadata. It never asks an LLM to invent descriptions, prices, parameters, or capabilities.

​
Indexing pipeline

Cataloging writes the document and queues embedding work after payment decisions. Request latency never waits for model inference. Workers claim jobs with PostgreSQL leases and fencing tokens, batch provider calls, retry with backoff, and dead-letter poisoned documents.

​
Canonical search text

The formatter uses declared fields only:
MCP fields include toolName and parameter descriptions from the existing inputSchema. Missing fields are omitted, not inferred.

​
Retrieval

  1. Apply structured filters before ranking.
  2. Retrieve PostgreSQL full-text candidates.
  3. Retrieve vector candidates from the active model generation.
  4. Fuse candidate ranks with reciprocal-rank fusion.
  5. Optionally rerank the fused head.
  6. Apply deterministic cursor and origin-diversity tie breaks.
The production formula is:
Only ranks enter the formula. PostgreSQL ts_rank_cd is not BM25, and cosine distance is never added to it as if both were comparable scores. The checked-in self-hosted profile uses lexical weight 0.7, semantic weight 0.3, rrf_k: 20, and candidate pools of 250. Search evaluation also measured a tuned development configuration, but that does not silently change the shipped profile. Operators should publish which configuration produced a benchmark result.

​
Model generations

Every generation records:
Only one generation is active for serving. A model, revision, dimension, pooling, or normalization change creates a new typed vector table and requires an explicit reindex. Vectors from different generations are never compared. The default local embedding identity is BAAI/bge-m3 through the pinned Xenova/bge-m3 ONNX artifact at 1024 dimensions. The Railway profile uses an OpenAI-compatible remote embedding endpoint and does not download model files. The checked-in model licence record is: The facilitator validates the returned vector dimension before storage. A provider whose terms or weights do not satisfy an operator’s redistribution policy must stay behind the remote-provider interface or be disabled.

​
Degradation ladder

Use /analytics/v1/search/status to inspect provider health, active generation, queue depth, coverage, and the exact degraded reason.

​
Query safety and provenance

Queries are normalized before PostgreSQL parsing. Empty or stopword-only input becomes a guaranteed no-match query. Seller descriptions are untrusted content: they are inert metadata and never become agent instructions. Search responses carry an x-search-session-id. A later resource fetch can use that identifier to record search-to-resource conversion. This signal is operator analytics, not a relevance judgment. See evaluation for measured model and retrieval results and configuration for provider settings.
Last modified on August 10, 2026