These packages give you a complete, modular RAG stack built around Qdrant, with hybrid retrieval that combines vector search, BM25 keyword search, and cross-encoder reranking. You'd adopt them to avoid assembling and integrating a dozen separate tools for document ingestion, chunking, retrieval, reranking, evaluation, and observability into a coherent pipeline. The most distinctive thing is that every component—from four chunking strategies to ablation studies to an MCP server with 41 tools—is a separate, independently installable package sharing core types and Zod schemas, so you can use only what you need while keeping everything type-safe and composable.
Zod schemas, TypeScript types, and enums for documents, chunks, retrieval results, evaluation samples, ablation configs, and benchmarking metrics that serve as the shared type foundation across the `@reaatech/hybrid-rag-*` ecosystem.
A CLI tool for hybrid RAG (Retrieval-Augmented Generation) workflows, providing commands for document ingestion, querying, evaluation, ablation studies, benchmarking, chunking preview, and MCP server startup against a Qdrant vector database.
A class that generates text embeddings through a provider-agnostic interface, currently supporting OpenAI models with built-in batch processing, rate limiting, and cost tracking.
An evaluation runner for hybrid RAG systems that provides standard IR metrics (Precision@K, Recall@K, NDCG@K, MAP, MRR), generation quality scoring, ablation studies with YAML-configurable variant testing, and latency/throughput/cost benchmarking. Exports `EvaluationRunner`, `AblationRunner`, and benchmarking functions.
A set of classes (`DocumentLoader`, `TextPreprocessor`, `DocumentValidator`) and a `chunkDocument` function for loading, preprocessing, validating, and chunking documents from PDF, Markdown, HTML, and plain text files, with four configurable chunking strategies (Fixed-Size, Semantic, Recursive, Sliding Window) and deterministic chunk IDs.
An MCP server that exposes 41+ tools for hybrid RAG (vector + BM25) operations, including retrieval, ingestion, evaluation, query analysis, session management, and agent integration. It provides a `createMCPServer` function that accepts a `RAGPipeline` instance and supports stdio, HTTP, and SSE transports.
A Pino-based structured logger and OpenTelemetry tracing/metrics collector for hybrid RAG pipelines, providing pre-built helpers for logging query lifecycles, recording span durations across retrieval stages, and collecting counters/histograms for queries, reranker calls, embeddings, and costs.
A single `RAGPipeline` class that orchestrates document ingestion and hybrid retrieval (vector + BM25) against a Qdrant vector store, with optional reranking via Cohere, Jina, OpenAI, or a local cross-encoder.
A Qdrant vector database adapter that wraps `@qdrant/js-client-rest` with collection management, batch upsert, vector search with automatic metadata filtering, and health checks, exported as a `QdrantClientWrapper` class.
A hybrid retrieval engine that combines Qdrant vector search, in-process BM25 keyword search, cross-encoder reranking, and configurable score fusion (RRF, weighted sum, normalized score) into a single `HybridRetriever` class.
