Download the PHP package cognesy/instructor-retrieval without Composer

InstructorPHP Retrieval

Vector storage, document indexing, semantic retrieval, and bounded RAG context for InstructorPHP. The package follows Symfony AI's composition model: one replaceable store contract is shared by direct vector persistence, the document processor/indexer, and retrieval. Instructor's facade and pending execution model remains the public orchestration surface.

Installation

The package bundles memory, pgvector, qdrant, typesense, meilisearch, weaviate, and milvus. The HTTP-backed drivers use Instructor's shared HTTP client; Pgvector requires ext-pdo and a configured PDO connection. Backend management is explicit through CanManageStore—constructing a store never creates or drops remote resources.

Configure bundled drivers

StoreConfig::metric selects the package-level metric; backend connection and schema values belong in options. StoreFactory resolves the configured name through the same replaceable registry used by the other InstructorPHP driver packages.

All drivers support bounded VectorQuery and reject ranked continuations. Their additional capabilities are deliberately backend-specific:

The required and driver-specific options are:

Typesense rejects Euclidean configuration. Meilisearch uses a userProvided embedder and supports cosine only; setup and every mutation wait for terminal task success before returning. Weaviate requires an endpoint with /v1/graphql enabled. Milvus intentionally does not expose CanScanDocuments: its REST offset cap cannot satisfy the package's large-corpus traversal contract.

Weaviate and Milvus expose squared L2 values on the wire; their adapters return the square root so SearchScore consistently represents the package's actual Euclidean distance. Vectors remain omitted unless requested with Projection.

Store precomputed embeddings

Embedding generation is not required when vectors already exist.

VectorDocument is deliberately generic. Text, source/chunk provenance, and even content itself are optional. Search results omit vectors unless the caller explicitly requests them with Projection.

Index text and retrieve by meaning

DocumentIndexer delegates to DocumentProcessor; the processor owns filtering, transformation, bounded batching, vectorization, persistence, and stale-chunk cleanup. Vectorizer is a thin adapter over Polyglot Embeddings.

Use separate document/query Vectorizer instances when a provider requires different task options, while keeping both roles in the same declared embedding space. Known dimension or embedding-space mismatches fail explicitly.

Compose RAG without hiding generation

Retrieval never invokes an answering model. ContextAssembler selects textual evidence under exact rendered byte, token, evidence-count, and excerpt limits; the application then passes that untrusted evidence to Inference or StructuredOutput.

The returned AssembledContext includes the exact citation map. Empty or content-free hits produce an explicit empty context; the application chooses whether to abstain, ask a follow-up question, or answer without evidence.

Traverse large corpora

Ranked retrieval always returns one bounded page. Corpus traversal is a separate scan capability with a forward-only cursor that releases previous pages.

Do not use ranked search as an export API. Drivers expose truthful capabilities for ranked continuation, scan, fetch, exact/approximate behavior, and ordering.

Agents

packages/agents depends on this package and provides UseRetrieval. It installs a stateless, bounded retrieval_search tool and optionally a retrieval_read tool. Reads require an application-supplied CanReadRetrievalEvidence implementation, which is the authorization boundary for execution, tenant, and collection scope. The generic AgentLoop has no vector-store dependency.

Multimodal content

The storage contract already accepts vectors for products, images, audio, or any other modality because a record does not require text. Applications can store externally generated multimodal embeddings and keep typed asset references in metadata. The bundled document processor, SemanticQueryVectorizer, and ContextAssembler are intentionally text-first in this release because Polyglot's current embeddings request is string-based. They do not silently flatten binary/image/audio inputs into text.

First-party multimodal vectorization should be added when Polyglot exposes a typed multimodal embeddings input. Multimodal RAG will then need a content-part evidence representation and modality-aware budgets; until then, hydrate an authorized asset reference in application code and pass it to the existing multimodal inference APIs.

Observability and safety

Retrieval and indexing dispatch lifecycle, store-attempt, and progress events through the shared events/logging stack. RetrievalTelemetryProjector maps a retrieval lifecycle to trace-ready data. Default events contain correlation IDs, driver/query types, counts, durations, and error classes—not vectors, credentials, query text, or retrieved content.

All HTTP-driver response bodies are capped before complete JSON decoding. Pgvector reads rows incrementally and closes PDO cursors explicitly. Store scans and retrieval tools remain bounded even when the backing collection is large.

Live backend tests

Integration tests are skipped unless their backend endpoint is configured:

Each live test owns a randomly named collection or index and exercises explicit setup, replacement, filtered query, projection, fetch, deletion, clear, and drop. Scan is additionally exercised where the driver advertises it.