txtai

Unified embeddings storage layer combining dense vector indexes (FAISS, Annoy, HNSW), sparse BM25 indexes, graph networks for relationship modeling, and SQL relational storage in a single queryable index. Supports multiple vector model backends (sentence transformers, local LLMs, API-based embeddings) with automatic quantization, persistence, and recovery. Implements co-location of vector, graph, and relational data enabling complex queries across all three modalities without separate systems.

Orchestrates retrieval-augmented generation by composing embeddings search, context ranking, prompt templating, and LLM inference into a configurable pipeline. Supports multiple LLM backends (OpenAI, Anthropic, Ollama, local transformers) with provider-agnostic prompt engineering. Implements context ranking strategies (BM25, semantic similarity, reranking models) to optimize retrieved context quality before passing to LLM, reducing hallucination and improving answer relevance.

Relational database layer enabling storage of structured metadata alongside embeddings and graphs. Supports multiple backends (SQLite, PostgreSQL, MySQL) with automatic schema creation. Enables SQL queries on metadata (filtering, aggregation) combined with semantic search. Implements full-text search on text columns and supports complex WHERE clauses for precise filtering.

Clustering layer enabling horizontal scaling of txtai across multiple machines. Implements index sharding (partitioning embeddings across nodes), request routing to appropriate shards, and result aggregation. Supports multiple sharding strategies (hash-based, range-based). Coordinates cluster state and handles node failures with automatic failover. Enables transparent scaling without application code changes.

Persistence layer enabling saving and loading of embeddings indexes to disk. Implements automatic snapshots at configurable intervals for disaster recovery. Supports incremental updates to avoid full index rewrite. Handles recovery from crashes with automatic index validation and repair. Enables reproducible results by persisting exact index state.

Declarative workflow engine that composes tasks (pipelines, agents, custom functions) into directed acyclic graphs (DAGs) defined in YAML configuration. Supports task dependencies, conditional branching, parallel execution, and scheduling via cron expressions. Implements task state management, error handling with retry logic, and result passing between tasks through a shared context object. Enables non-technical users to define complex AI workflows without code.

Agent framework enabling autonomous task execution through iterative reasoning loops (think → act → observe). Agents have access to tool registry (function calling) with native bindings for common APIs and custom tools. Implements agent teams for collaborative multi-agent workflows where agents delegate tasks, share context, and coordinate toward goals. Uses LLM reasoning for tool selection and execution planning with built-in safety guardrails and execution limits.

Extensible pipeline architecture supporting specialized processing chains for different modalities: text (NLP, summarization), audio (transcription, speech-to-text), image (OCR, classification, object detection), and data (ETL, transformation). Each pipeline type implements a standard interface enabling composition into larger workflows. Pipelines are configured declaratively and can be chained together with automatic type conversion between modalities.

Application layer exposing all txtai components (embeddings, pipelines, workflows, agents) via HTTP REST endpoints and OpenAI-compatible chat/completion APIs. Implements Model Context Protocol (MCP) for integration with Claude and other MCP-compatible clients. Handles request routing, authentication, clustering/sharding coordination, and response serialization. Enables deployment of txtai applications as microservices without code changes.

Semantic search capability combining dense vector similarity (learned embeddings) with sparse keyword matching (BM25) in single query. Implements multiple ranking strategies: semantic similarity scoring, BM25 keyword matching, and optional neural reranking models. Supports filtering by metadata, date ranges, and custom predicates. Returns ranked results with relevance scores and supports pagination for large result sets.

Embedding inference engine supporting multiple model sources: sentence-transformers, local transformers, and API-based providers (OpenAI, Hugging Face). Implements automatic quantization (int8, float16) to reduce model size and inference latency. Caches embeddings to avoid recomputation and supports batch inference for efficiency. Abstracts model provider differences enabling seamless switching between local and API-based embeddings.

Application class that reads YAML configuration files and instantiates all txtai components (embeddings, pipelines, workflows, agents) with dependency injection. Manages component lifecycle (initialization, shutdown, persistence). Supports environment variable substitution in YAML for deployment flexibility. Enables reproducible application setup and deployment without code changes.

Graph database layer enabling storage and traversal of relationships between entities. Supports directed and undirected edges with properties, enabling knowledge graph construction. Implements graph algorithms (shortest path, community detection, centrality) for relationship analysis. Integrates with embeddings database enabling hybrid queries combining semantic similarity with graph traversal.