RAG and Retrieval

20 resources

• qdrant/qdrant

  A vector database for similarity search with structured filtering and multiple representations per point.

  qdrant · Local server or managed service.

  Compatibility

  Docker, Python, REST, gRPC

  Components

  Vector database, Metadata filtering, Hybrid retrieval

  Setup notes

  Use named vectors to keep dense embeddings and sparse features associated with the same chunk ID. Filter on fields such as collection, language, or tenant when the same service indexes multiple sources.

  Run Qdrant locally in its official Docker container and connect through its REST or gRPC API. Choose collection dimensions to match the embedding model, and save the original chunk text or its document pointer alongside the vector.

• Milvus Vector Database

  An open-source vector database with configurable indexes and scalar fields for filtering retrieval.

  Milvus Project · Local library or server. Managed options are separate.

  Compatibility

  Python, Docker, PyMilvus

  Components

  Vector database, Metadata filtering, Vector indexes

• Weaviate Vector Database

  A vector database that can combine embedding similarity, keyword relevance, and property filters.

  Weaviate · Local server or managed service. Some vectorizer integrations require credentials.

  Compatibility

  Docker, Python, REST, GraphQL

  Components

  Vector database, Keyword search, Hybrid retrieval

• Chroma

  An embedding store with a local Python client, collections, and document metadata filters.

  Chroma · Local Python client or server.

  Compatibility

  Python, JavaScript, Docker

  Components

  Vector store, Metadata filtering, Embedding storage

• LanceDB

  An embedded vector database for storing, filtering, and retrieving chunks from application data.

  LanceDB · Local embedded database or managed service.

  Compatibility

  Python, TypeScript, Rust

  Components

  Embedded vector store, Full-text search, Hybrid retrieval

• pgvector for PostgreSQL

  A PostgreSQL extension that adds vector columns, similarity operators, and approximate search indexes.

  pgvector contributors · Local PostgreSQL extension. Requires a PostgreSQL installation that supports the extension.

  Compatibility

  PostgreSQL, SQL

  Components

  Vector database extension, SQL filtering, Approximate nearest neighbors

• faiss-cpu

  A library for exact and approximate nearest-neighbor search over dense numeric vectors.

  Meta AI Research · Local C++ or Python library. CPU and supported GPU builds are available.

  Compatibility

  Python, C++, CPU, GPU

  Components

  Vector indexing, Approximate nearest neighbors, Embedding search

  Setup notes

  Use a simple flat index as the correctness baseline before choosing an approximate index. Keep the passage text and application metadata in a separate store keyed by your FAISS IDs.

  Benchmark recall at a fixed top-k, query latency, and index memory on the same embeddings. Choose a distance and normalization strategy that matches the embedding model and use it throughout the comparison.

• Vespa

  A search platform with tensor fields, lexical retrieval, and configurable multi-stage ranking.

  Vespa · Local deployment or managed service.

  Compatibility

  Docker, REST, Vespa schemas

  Components

  Search engine, Tensor ranking, Hybrid retrieval

• Haystack

  A Python framework for assembling document indexing, retrieval, reranking, and answer pipelines.

  deepset · Local Python package. Hosted model and search integrations can require credentials.

  Compatibility

  Python, Local inference, Hosted integrations

  Components

  RAG pipeline, Document indexing, Retrieval orchestration

• LlamaIndex

  A Python framework for loading data, creating document nodes, indexing them, and composing retrieval queries.

  LlamaIndex · Local Python framework. Provider integrations can require credentials.

  Compatibility

  Python, Local inference, Hosted integrations

  Components

  Document ingestion, RAG pipeline, Retrieval orchestration

• docling

  A document conversion toolkit that preserves structure such as sections, reading order, and tables.

  docling-project · Local Python toolkit. Initial model assets may need downloading.

  Compatibility

  Python, PDF, Office documents

  Components

  Document parsing, OCR, Table extraction, Chunk preparation

  Setup notes

  Convert a small representative set first and inspect reading order, headings, and tables. These structures can become chunk boundaries or metadata used to reconstruct citations.

  Run Docling locally and keep source references as you export Markdown or structured output. OCR and conversion settings can affect both runtime and the extraction quality of scanned pages.

• unstructured

  A Python document-processing library that partitions supported files into labeled content elements.

  unstructured-io · Local Python library. Some file types require additional system packages.

  Compatibility

  Python, PDF, HTML, Office documents

  Components

  Document parsing, Element extraction, Chunk preparation

  Setup notes

  Preserve element type and page metadata during partitioning so the index can retain document context. Compare outputs on PDFs that contain tables, multi-column layouts, and scanned pages before processing the full collection.

  Choose a local partitioning strategy according to document complexity and installed dependencies. Save chunk IDs and source locations before sending the prepared text to a vector store.

• sentence-transformers

  A Python library for local text embeddings, semantic retrieval, and cross-encoder reranking.

  huggingface · Local Python inference. Model weights may need downloading.

  Compatibility

  Python, PyTorch, Hugging Face models

  Components

  Embeddings, Cross-encoder reranking, Semantic retrieval

  Setup notes

  Use a bi-encoder to search a corpus efficiently, then apply a cross-encoder only to the candidate passages. The second stage scores each query and passage together.

  Keep the first-stage candidate cutoff fixed while comparing rerankers. Measure the ranking improvement and latency separately, and check the selected model license before deployment.

• flashrank

  A Python reranking library with compact models for reordering passages from an existing retriever.

  Prithivi Da · Local Python inference. Initial model files may need downloading.

  Compatibility

  Python, ONNX

  Components

  Passage reranking, Local model inference

  Setup notes

  Run the first-stage retriever as usual, then supply its passage text and query to FlashRank. Use the returned relevance scores to order the passages before building the generator context.

  Compare the same candidate set with and without reranking. Record the model name, candidate cutoff, and time spent reranking so you can measure the tradeoff on your own corpus.

• Hybrid retrieval with PostgreSQL and pgvector

  A database-side recipe that combines PostgreSQL text search with vector similarity using rank fusion.

  Supabase · SQL recipe for PostgreSQL with pgvector. Embeddings are supplied separately.

  Compatibility

  PostgreSQL, pgvector, SQL

  Components

  Hybrid retrieval, PostgreSQL full-text search, Reciprocal rank fusion

• Dense and sparse retrieval with Qdrant Query API

  A retrieval recipe for merging candidate lists from multiple vector representations in one request.

  Qdrant · Local or managed Qdrant with compatible Query API and embedding models.

  Compatibility

  Qdrant, Python, REST

  Components

  Hybrid retrieval, Sparse vectors, Reciprocal rank fusion

• Keyword and semantic search with OpenSearch

  An OpenSearch workflow for combining lexical and semantic results through a search pipeline.

  OpenSearch Project · Requires an OpenSearch deployment and a configured embedding model.

  Compatibility

  OpenSearch, REST, JSON

  Components

  Hybrid retrieval, Search pipeline, Score normalization

• Retrieve then rerank with Sentence Transformers

  A local retrieval recipe that searches with a fast first stage and rescores candidate pairs with a cross-encoder.

  Sentence Transformers contributors · Local Python recipe. Download the selected model weights before offline runs.

  Compatibility

  Python, Sentence Transformers, PyTorch

  Components

  Two-stage retrieval, Cross-encoder reranking, Embedding search

• Structure-aware chunking with Docling

  A document-preparation recipe for balancing document structure with an embedding model’s token budget.

  Docling Project · Local Python recipe. Tokenizer or model artifacts may need downloading.

  Compatibility

  Python, Docling, PDF

  Components

  Structure-aware chunking, Document parsing, Token budgets

• Hybrid vector and full-text search in LanceDB

  A table-level retrieval recipe for combining vector similarity and full-text results.

  LanceDB · Local LanceDB recipe with full-text indexing and an embedding model.

  Compatibility

  Python, LanceDB

  Components

  Hybrid retrieval, Full-text indexing, Vector search