RAG · Search

This repo open-sources a full Chinese textbook on AI agent engineering—complete with Markdown source, compiled PDF, and runnable Python demos for every chapter.

wigolo gives AI coding agents a local, keyless web research layer that runs entirely on your machine and bills nothing per query.

Moss exists because calling out to a remote vector database adds 200–500 ms of latency—enough to kill a real-time conversation—so it runs embedding and search inside your process instead.

Heym exists because serious AI automation shouldn't require a sales call to access observability, evals, or multi-agent orchestration.

Vector search shouldn't force you to choose between PostgreSQL and DuckDB.

Macro exists because context dies when your email, chat, docs, and tasks live in separate apps, so it links them into one queryable workspace with shared AI memory.

Omnigraph exists to let fleets of agents collaborate on a shared knowledge graph using git-style branches and merges, keeping the data on S3-compatible object storage.

turbovec exists so you can index embeddings immediately—no training, no tuning, no rebuilds—and search them faster than FAISS in a fraction of the RAM.

Turns URLs into clean markdown and JSON so AI agents don't have to parse HTML soup.

It exists because RAG shouldn't be the only way to turn a file dump into usable knowledge.

It exists because builders wanted a Claude Cowork-style agent that keeps work context local, encrypted, and auditable instead of shipping memory to someone else’s server.

PixelRAG renders documents into screenshot tiles and retrieves them visually, preserving tables and layout that HTML parsers strip away.

Built for Korean offices drowning in HWP and Office files, this Rust/Tauri app indexes full document text locally so you can find files by what they contain, not what they’re named.

MEMANTO treats memory as an active agent, not a passive vector dump, with typed storage and instant retrieval.

LycheeMemory gives LLM agents persistent, structured long-term memory across runtimes without the usual token bloat or graph-database baggage.

memsearch gives Claude Code, Codex CLI, and other agents a unified, persistent memory layer backed by Markdown and Milvus.

A pipeline that turns messy PDFs and slides into structured, navigable memory for AI agents instead of flat text shards.

Second Brain is a self-hosted semantic memory layer that lets every AI tool you use—Claude, ChatGPT, Cursor, or any MCP client—recall the same persistent context without you repeating yourself.

It packages Karpathy's LLM wiki concept as an Agent Skill so your coding agent can maintain a durable knowledge base with citations instead of re-deriving answers from raw sources on every query.

Scout exists because chunking your entire company into a vector database is a brittle way to build a corporate memory.