Skip to main content
v2.12.0 adds vector retrieval and structured decisions to InstructorPHP, expands exact-model metadata, and brings new Echo and RESPAN integrations. It also includes an embedding configuration cutover; review the upgrade notes before updating.

​
Retrieval and agent knowledge access

  • The new cognesy/instructor-retrieval package provides document processing, vector storage, semantic search, bounded context assembly, and citations. Bundled stores include in-memory, pgvector, Qdrant, Typesense, Meilisearch, Weaviate, and Milvus.
  • Retrieval composes with Polyglot Embeddings and leaves answer generation to the application, so callers control how evidence enters inference requests.
  • Agents can opt into bounded search and evidence-reading tools. Query, result, context, and read sizes are policy-controlled. Evidence reads pass through an application-owned authorization boundary.

​
Structured decisions and model metadata

  • Polyglot adds a typed Decision runtime for Noul, Choice, and Score questions. Adapters cover Classifier.dev, Jeff, Laya, TypeSafe, and RESPAN.
  • Noul answers can carry positive, negative, and unknown probabilities. RESPAN supports Noul questions and reports access failures separately from request validation errors.
  • Exact Embedding and Decision model catalogs expose reviewed capabilities and limits, with pricing where available. Examples show model metadata and cost estimates.

​
Provider and integration updates

  • The Echo inference driver supports Fulcrum’s persona-based writing API and requires a non-empty persona option on each request or preset.
  • Embedding request model overrides now select the effective model for catalog lookup and provider request construction. Missing provider token counts remain unknown instead of being converted to zero.
  • The cURL transport negotiates HTTP/2 over TLS by default and handles interim response headers before the final response.

​
Upgrade Notes

  • Embedding configuration no longer accepts dimensions, defaultDimensions, or maxInputs in EmbeddingsConfig, DSNs, presets, or Laravel and Symfony embedding connection configuration. Set a requested dimension in request options, for example withOptions(['dimensions' => 256]). Inject an Embedding model catalog into EmbeddingsRuntime when catalog-based input-limit checks are required.
  • Embedding usage counts and totals can now be null when a provider omits them, and embedding cost calculation can return null. Handle unknown usage explicitly. Use EmbeddingsUsage::zero() for a known zero total; none() represents unknown usage.
  • Add ECHO_API_KEY or RESPAN_API_KEY only for applications that use the corresponding provider.