@convex-dev/rag

Automatically converts documents into dense vector embeddings using configurable embedding models (OpenAI, Anthropic, or local alternatives) and persists them in Convex's serverless database with metadata indexing. The system handles chunking strategies, batch processing, and incremental updates without requiring external vector databases like Pinecone or Weaviate.

Executes vector similarity queries against stored embeddings using cosine distance, dot product, or Euclidean distance metrics. Queries are performed via Convex functions that compute similarity scores between a query embedding and all stored document embeddings, returning ranked results with configurable result limits and filtering predicates applied before or after similarity computation.

Automatically splits long documents into semantically coherent chunks using configurable strategies (character-based, token-based, or recursive with overlap). The framework handles chunk size limits, overlap windows to preserve context, and metadata propagation so each chunk retains references to the original document and its position, enabling retrieval of full context during RAG synthesis.

Provides a pluggable interface for embedding generation supporting OpenAI, Anthropic, and local/self-hosted models through a unified API. The framework abstracts provider-specific details (API endpoints, authentication, request/response formats) so developers can switch embedding models without changing application code, and handles retries, rate limiting, and error recovery transparently.

Provides utilities to retrieve relevant documents from semantic search results and format them as context for LLM prompts, handling token budgeting, context window management, and integration with LLM APIs (OpenAI, Anthropic, etc.). The framework manages the retrieval-augmented generation loop: query → embed → search → retrieve → format context → call LLM → return answer.

Automatically detects document changes and re-embeds only modified documents rather than rebuilding the entire index. The system tracks document versions, timestamps, and change hashes to identify which documents need re-embedding, and handles concurrent updates safely within Convex's transactional guarantees without requiring manual index invalidation or rebuild triggers.

Processes multiple documents in batches through the embedding API, handling rate limiting, transient failures, and partial failures gracefully. The framework groups documents into optimal batch sizes for the embedding provider, implements exponential backoff retry logic, and tracks which documents succeeded/failed so applications can retry failed embeddings without re-processing successful ones.

Combines semantic similarity search with metadata-based filtering and optional keyword matching to refine results. The framework applies metadata predicates (e.g., 'category=finance AND date>2024') before or after similarity computation, and can optionally incorporate keyword/BM25 scoring alongside vector similarity for hybrid ranking that balances semantic relevance with exact term matches.