Retrieval

The job

A model only knows what is in its prompt and its training. Retrieval tools fetch material at the moment of a question, whether that is source code, library documentation, web pages or company files. Agent builders reach for them when answers go stale, or when the material is too large to paste in.

How they differ

One kind maps a codebase so a coding agent reads less. Codegraph and graphify build a knowledge graph of a repository, and zvec-grep combines exact text, keyword and vector search over a local index. Openwiki has an agent write a linked Markdown wiki. These may suit developers in large repositories.

A second kind supplies library documentation. Context7 pulls version-specific docs into the agent's context. It may suit people whose agent gives out-of-date API answers.

A third kind answers questions over your own documents. Ragflow parses mixed files and cites sources. Private-gpt is an API server in front of a local model server. Rocketride-server is a pipeline engine whose search pipelines can include a vector database. These suit teams who can host a service.

A fourth kind gathers web pages. Crawl4ai and firecrawl turn pages into Markdown or structured data for agents, and both offer a hosted API.

A fifth kind stores facts for agents to query. Utopia sends agent writes to a person for review, gbrain lets facts be withdrawn, semantica builds knowledge graphs in Python, and xerj adds a memory store to search.

Licences

On 10 November 2026, GitHub detected Apache-2.0 or MIT on most of the tools here. Firecrawl carried AGPL-3.0. A licence can change, and each tool's own page gives the latest reading with its date.

What to check first

Where content goes matters first. Firecrawl says its quickstart calls a hosted API with a key, so requests pass through its service. Crawl4ai says its hosted API routes target URLs through its own cloud, on credits. Context7 disclaims the accuracy and security of the community docs it serves. Graphify says documents, images and video reach whichever model your assistant or key points to. Gbrain says configured cloud providers can receive text, and warns of an unrelated npm package that can replace its binary.

Keys and files need attention too. Openwiki says installing an integration edits user-level files, and its connectors can store OAuth tokens for Gmail, Slack and X. Ragflow and rocketride-server need model keys. Private-gpt lists code execution and web fetch among its built-in tools and says to check how they are isolated.

Codegraph pipes a shell script on one install route and has a telemetry section. Utopia, xerj and semantica carry early-release notes, and xerj uses an insecure flag in examples. Paid or hosted parts are named by ragflow, crawl4ai, firecrawl and gbrain.