AI · RAG, Embeddings & Vectors

The vector index sits beside your documents. Unlike a separate vector database, it removes the copy-and-sync burden; full support is on Atlas, and the feature is newer in self-managed MongoDB.

The most widely used family of open embedding models. Compared with paid embedding APIs it sits close on the MTEB leaderboard, and most releases are MIT — commercial use is unblocked and data stays on your machine.

A vector database that runs embedded in your application with no server process — usable straight after pip install. Unlike server-based ones such as Qdrant or Milvus it is among the quickest for a prototype; at production scale you may hand off to something else.

Represents 100+ languages in one vector space, so a Turkish query can retrieve an English document. Unlike OpenAI embeddings it takes separate input types for query and document — the distinction that lifts match quality.

Combines BM25 keyword search with vector search in a single query and fuses the results with RRF. Unlike pure vector databases, it has exactly that keyword half.

Walks a codebase via AST analysis and folds in docs, SQL schemas and PDFs to build a queryable graph. Unlike plain text search, it hands a coding agent the relationships — who calls whom, which table is used where — as context.

Extracts entities and relations into a knowledge graph, then builds community summaries. Unlike classic chunk-retrieval RAG, it can answer corpus-wide questions such as “what are the main themes here”.

You wire components together explicitly, so what happens at each stage stays readable. Unlike LangChain’s implicit flow it is transparent — a visible win when debugging.

Multimodal embedding and search infrastructure that makes documents and images searchable by vector. Unlike single-modality embedding APIs it places text and images in one space, and it also offers a reader that turns web pages into model-ready text, plus rerankers.

The concept page separating short-term memory (in-thread state) from long-term (a cross-thread store), and when each applies. Unlike general writing on agent memory, it ties the distinction to the concrete LangGraph API.

The productised form of the MemGPT research. Unlike flat vector memory, it manages the context window like an operating system, keeping hot state in the window and paging the rest to external storage.

Makes each stage — loading, chunking, indexing, querying — independently swappable. Not a general-purpose framework: it is built for someone tuning a retrieval pipeline.

An end-to-end vector search engine that handles everything from embedding generation to search. Instead of standing up a separate embedding service and vector database, you hand it raw text or images and search directly.

Stores facts extracted from conversation, updates contradictions and drops the irrelevant. A layer that decides what is worth remembering instead of accumulating raw history.

The self-hostable version of the Mem0 memory layer. Instead of the hosted Mem0 platform you keep the data yourself; the Embedchain project was merged in here too, so older Embedchain links land here.

A distributed vector database separating storage from compute, built for billion-scale collections. Compared with Qdrant it targets larger scale; the distributed setup wants Kubernetes, while Milvus Lite and a single-node build cover small projects.

Relationships are first-class data, not foreign keys. Multi-hop traversals in Cypher are both shorter and faster than the nested JOINs the same question needs in SQL.

Reference for the text-embedding-3 family — dimensions, pricing, usage. Compared with open embedding models there is nothing to operate, and Matryoshka support lets you truncate vector dimensions on demand to cut storage cost.

An extension adding a vector type plus HNSW and IVFFlat indexes to PostgreSQL. Unlike a separate vector database, it keeps vectors in the same transaction as your relational data.

A fully managed vector database whose serverless tier removes capacity planning. Compared with self-run options such as Qdrant or pgvector it has the lowest operational burden, at the cost of control over where the data sits.

With JSONB, full-text search and pgvector it now does the work of three separate systems. Instead of running a document store, a search engine and a vector database separately, it simplifies early-stage architecture considerably.

A vector database written in Rust that executes heavy metadata filtering alongside the vector search. Unlike Pinecone it is open source and self-hostable; its low memory footprint lets a small VPS carry a serious collection.

A metrics library for RAG pipelines, including metrics that work without reference answers. Instead of evaluating by reading samples by hand, it can generate a synthetic test set so evaluation runs in CI.

Built on deep document parsing: it recognises complex PDF layouts, tables and figures before chunking. Not a library but a full service you stand up with Docker.

A memory-first agent harness. Unlike agents that start every session from scratch, it centres on learning across sessions and refining its own configuration.

An in-memory data structure store, used on the AI side for session state, rate-limit counters and response caching. Compared with a separate vector database, its advantage is that vector search is added to the cache you already run.

The standard Python library for producing embeddings. Unlike embedding APIs the model runs locally — data stays put and there is no API bill; in exchange, GPU and memory management are yours.

Measures the RAG triad — context relevance, groundedness, answer relevance. Unlike evaluations that give one overall score it measures the three separately, which tells you whether the failure sits in retrieval or generation.

Reduces PDF, DOCX, HTML, email and slides to a shared document-element model. Unlike format-specific parsers it brings everything into one structure — and this first step is where RAG pipelines lose the most time and break most often.

Open-source AI answering engine (formerly Perplexica) that pairs web search with local or cloud LLMs to produce cited answers; runs self-hosted with local models unlike commercial Perplexity.

The embedding service inside GCP. Compared with another provider’s embedding API, if your data sits in BigQuery, generating vectors within the same boundary makes sense for egress cost and data residency.

Embedding models trained separately for law, code and finance; it joined MongoDB in 2025. Unlike general-purpose models it is domain-trained and claims to beat them in those domains — a claim to test on your own data.

An open-source vector database. Unlike databases that only store vectors, it can host embedding generation as a module, so you write raw text without a separate embedding step; its query interface is GraphQL.

Maintains a temporal knowledge graph for agents, recording when a fact held and when it was invalidated. Unlike flat vector memory, it can make exactly that distinction.