hybrid-rag-qdrant · packages

Provides a shared library of TypeScript interfaces and Zod schemas for defining documents, retrieval results, and evaluation metrics within a hybrid RAG pipeline. It serves as the type-safe foundation for the `@reaatech/hybrid-rag` ecosystem, requiring only `zod` as a runtime dependency.

Provides a command-line interface for managing hybrid RAG pipelines, including document ingestion, querying, evaluation, and benchmarking. It requires a Qdrant instance and supports running as an MCP server for integration with LLM-based applications.

Generates vector embeddings through a unified interface for OpenAI, Vertex AI, and local models, featuring built-in batching, rate limiting, and cost tracking. It provides an `EmbeddingService` class that returns standardized objects containing the vector, token usage, and calculated cost.

Evaluates hybrid RAG systems by providing classes for running IR metrics, ablation studies, and performance benchmarking. It exposes runner classes and utility functions that accept custom query callbacks to measure retrieval accuracy, generation quality, latency, and cost.

Provides a suite of classes and functions for loading, normalizing, validating, and chunking documents into formats suitable for RAG pipelines. It supports four distinct chunking strategies and generates deterministic IDs for PDF, Markdown, HTML, and plain text files.

Exposes over 40 Model Context Protocol (MCP) tools for managing RAG lifecycles, including retrieval, ingestion, evaluation, and observability. It provides a `createMCPServer` function that wraps a `RAGPipeline` instance and supports stdio, HTTP, or SSE transport layers.

Provides structured logging, OpenTelemetry tracing, and metrics collection specifically for hybrid RAG pipelines. It exports a set of utility functions and managers that wrap Pino and OpenTelemetry to track query lifecycles, retrieval performance, and system costs.

Provides a unified `RAGPipeline` class for orchestrating document ingestion, hybrid vector and BM25 retrieval, and reranking. It requires a Qdrant instance and integrates with various embedding and reranking providers to manage the end-to-end RAG lifecycle.

Provides a wrapper class for the Qdrant REST client that simplifies collection management, batch upserting, and metadata-filtered vector searches. It acts as an abstraction layer over `@qdrant/js-client-rest` to streamline common RAG operations.

Orchestrates hybrid RAG pipelines by combining Qdrant-based vector search, in-process BM25 keyword search, and cross-encoder reranking. It provides a `HybridRetriever` class that manages score normalization and fusion strategies like RRF or weighted sum to return a unified list of search results.