Pinecone MCP Server

Inserts or updates vectors in a Pinecone index with associated metadata and IDs through MCP tool interface. Implements batch upsert operations that accept vector embeddings (float arrays), unique identifiers, and arbitrary JSON metadata, routing them to the Pinecone API with automatic connection pooling and error handling. Supports sparse-dense vector formats for hybrid search scenarios.

Queries a Pinecone index using vector similarity search with optional metadata filtering and result ranking. Accepts a query vector (or raw text that gets embedded), performs approximate nearest neighbor search using Pinecone's indexing structure (HNSW or IVF), and returns top-k results with similarity scores. Supports metadata filter expressions to constrain results to specific subsets (e.g., documents from a date range or category).

Creates, deletes, and lists Pinecone indexes and namespaces through MCP tools. Manages index configuration (dimension, metric type, pod type) and namespace isolation for multi-tenant or multi-project scenarios. Provides introspection into index statistics (vector count, dimension, metric) and namespace-level operations without direct API calls.

Deletes vectors from a Pinecone index using metadata filter expressions or by explicit ID. Supports bulk deletion by filter (e.g., delete all vectors with timestamp < X) or individual deletion by vector ID. Operates at namespace level and returns count of deleted vectors.

Isolates all vector operations (upsert, query, delete) to specific namespaces within a Pinecone index. Namespaces provide logical partitioning of vectors without requiring separate indexes, enabling multi-tenant or multi-project scenarios. Each operation accepts an optional namespace parameter that routes to the correct partition.

Supports hybrid search combining sparse vectors (keyword/BM25 style) and dense vectors (semantic embeddings) in a single query. Accepts both sparse and dense vector representations, performs weighted combination of results, and returns unified ranked results. Enables keyword-aware semantic search without separate keyword index.

Executes multiple vector queries in a single MCP call and aggregates results with optional deduplication and ranking. Accepts array of query vectors or text queries, performs parallel similarity search for each, and returns combined ranked results. Useful for multi-query retrieval patterns (e.g., query expansion, multi-hop reasoning).

Retrieves index metadata including vector dimension, similarity metric (cosine/euclidean/dotproduct), vector count, and index status. Provides runtime introspection for agents to validate query vectors and understand index configuration without external documentation.

Implements automatic retry logic for transient Pinecone API failures (rate limits, temporary outages) with exponential backoff, and provides detailed error messages for permanent failures (invalid API key, dimension mismatch). Handles MCP-specific error responses that include error codes, messages, and recovery suggestions. Reduces noise from transient failures while surfacing actionable errors.

Manages Pinecone API authentication through environment variables or configuration files, supporting secure credential storage without embedding keys in code. Implements credential validation at startup and automatic re-authentication if tokens expire. Supports both API key and OAuth-based authentication (if Pinecone supports it), with fallback mechanisms for missing credentials.