pinecone-client

Executes approximate nearest neighbor (ANN) search over dense vector embeddings using optimized indexing algorithms (tree-based or graph-based structures like HNSW), returning top-K results filtered by JSON metadata predicates. The client sends a query vector and optional filter constraints to the Pinecone managed service, which applies filtering before or after ANN traversal depending on selectivity, returning ranked results with scores and metadata in real-time (<100ms latency for typical workloads).

Executes full-text search using sparse vector representations (token-based, typically BM25-weighted) to find lexically similar documents, complementing dense semantic search. The client sends sparse vectors (token IDs with weights) to Pinecone, which applies inverted index lookups and BM25 ranking, enabling hybrid search when combined with dense results. Sparse vectors are more interpretable than dense embeddings and excel at exact keyword matching.

Lists vector IDs in an index or namespace, enabling pagination, auditing, or bulk operations. The client requests a list of IDs (optionally filtered by namespace or prefix); Pinecone returns paginated results. This is useful for understanding index contents or implementing cursor-based retrieval.

Authenticates client requests using API keys issued by Pinecone account setup. The client includes the API key in requests (via header or constructor parameter); Pinecone validates the key and authorizes operations. This is a simple, stateless authentication model suitable for server-to-server communication.

Deploys Pinecone indices in specific cloud regions (AWS, GCP, Azure) and availability zones, enabling data residency compliance and latency optimization. The client connects to indices in the selected region; Pinecone handles replication and failover within that region. This is configured at index creation time, not per-query.

Creates backups of vector indices and restores them to recover from data loss or enable point-in-time recovery. Pinecone manages backups automatically or on-demand; the client can trigger restore operations to recover a previous index state. Backup and restore are asynchronous operations.

Executes simultaneous sparse (lexical) and dense (semantic) vector search in a single query, combining results via weighted fusion (e.g., reciprocal rank fusion or linear combination of scores). The client sends both sparse and dense vectors to Pinecone, which performs parallel ANN and inverted index lookups, then merges ranked results using configurable fusion strategies. This enables retrieval systems that benefit from both keyword precision and semantic understanding.

Inserts or updates vectors with associated metadata in real-time, automatically indexing them for immediate search availability. The client sends upsert requests (vector ID, dense/sparse vector, metadata JSON) to Pinecone, which applies the vector to the ANN index and metadata to the filter index within milliseconds. Upserted vectors are queryable immediately without batch reindexing, enabling dynamic knowledge base updates in RAG systems.

Partitions vector data within a single index into isolated namespaces, enabling multi-tenant deployments where each tenant's vectors and metadata are logically separated. The client specifies a namespace string in query and upsert operations; Pinecone enforces isolation at the storage and query layers, ensuring queries in namespace 'tenant-A' never return results from 'tenant-B'. Namespaces share the same index infrastructure but maintain separate vector spaces.

Ingests large volumes of vectors and metadata from cloud object storage (S3, GCS) in batch, avoiding the need to stream individual upserts through the client. Pinecone reads vector files directly from object storage, parses them (format unspecified in deprecated docs), and indexes them in bulk. This is more efficient than client-side upsert loops for large-scale data migrations or initial index population.

Converts raw text directly to vectors using Pinecone-hosted embedding models (e.g., OpenAI, Cohere) without requiring external embedding infrastructure. The client sends text strings to Pinecone, which applies the configured embedding model and returns dense vectors. This eliminates the need to manage separate embedding services or pre-compute embeddings offline.

Reranks or filters search results after retrieval based on metadata attributes, enabling precision refinement beyond ANN scoring. The client can apply custom reranking logic (e.g., boost results with specific metadata values, sort by timestamp) to post-process Pinecone results. This is useful for business logic that cannot be expressed as pre-query filters (e.g., 'boost recent documents by 20%').

Retrieves specific vectors and metadata by their IDs without performing a search. The client sends a list of vector IDs to Pinecone, which returns the corresponding vectors and metadata. This is useful for retrieving known records or validating that vectors exist in the index.

Removes vectors from the index by vector ID or by metadata filter criteria. The client sends delete requests specifying either exact IDs or filter predicates; Pinecone removes matching vectors from both the ANN index and metadata index. Deleted vectors are immediately unavailable for search.