About TTalk 言几
A software and data company built around large language models.
The name: 言几
TTalk is short for [:strong "The Talk"]; the Chinese name is [:strong "言几"](yán jī).
- 言
- Yán (言) means speech, language — the Talk in TTalk, and the language in large language models.
- 几
- Jī (几, written 幾 in classical Chinese) is the faint first sign of things about to move. From the Book of Changes: “The yi is that by which the sages fathom the deepest and study the slightest signs.”
夫易,圣人之所以极深而研几也。
Together: seeing the sign in language, seizing the moment in the subtle. Yán is the software that lets models converse; jī is the craft of pulling signal out of data and retrieval.
What we do
Software products
Desktop apps that put large language models in ordinary people's hands: organizing reading material and writing long fiction. Everything lives as ordinary files on your own computer.
- Guanlan: An AI research assistant that lives on your computer. Point it at a folder to ask questions, organize cards, draw link graphs, build slides.
- Jinshu: An AI partner for long-form fiction. Characters, foreshadowing, plotlines, what the reader already knows — all remembered, with self-checks after every chapter.
- Coxswain: An Emacs-style editor rebuilt in Clojure: Emacs keybindings, tree-sitter highlighting, LSP, project tree, Git, REPL, and themes. Free download.
Data products
LLM applications live on retrieval: full text, vectors, hybrid fusion, plus transactions and graphs. We write our own storage and retrieval engines and embed them in the app — no extra service to run.
- libbitcask: A Bitcask append-only KV store in C++23, layered with BM25 full text, vector nearest-neighbor, and RRF hybrid search. A C++ library to use directly, plus a C ABI for bindings in other languages.
- Bitcask: libbitcask brought into Erlang/OTP via a NIF: KV, full text, vectors, hybrid search, and structured queries — plus transactions, idempotency keys, graph processing, and local embeddings.
Frameworks
The frameworks we built for our own products, cleaned up and open sourced: AI agent frameworks for Erlang/OTP and Clojure, template and Markdown rendering, and a server-rendered web framework.
- BeamAI: Build AI agents on Erlang/OTP: a unified LLM client across six vendors, ChatClient and tool calling, onion-style filters, and a ReAct agent with conversation memory.
- BeamAI Extra: Extensions on the BeamAI core: file / shell / human-interaction tools and middleware, RAG, an MCP client and server, the A2A protocol, and an AG-UI event-stream server.
- BeamAI Render: Mustache and Jinja2 templates compile to Erlang modules at build time — rendering is one function call — plus a Markdown engine ported from markdig with 30 extensions. Zero dependencies.
- aihtml: Assemble pages from Erlang functions with complete server-side HTML; clicks are answered statelessly by Erlang functions and server pushes ride SSE. 112 components, four theme axes, SEO-friendly.
- clj-agent: Build AI agents in Clojure: deftool for declaring tools, ChatClient and filter middleware, five hook types, a SimpleAgent that can pause for approval, and adapters for a dozen-plus model vendors.
How we work
- Your files belong to you: Your material and drafts are ordinary files on your computer — open them in any other program, back them up whenever you like. Guanlan never modifies a single original file.
- Runs on your own computer: No account required; searching your material happens locally. AI conversations can use DeepSeek, Qwen, Kimi and other online services with your own account.
- Changes pass through you: AI never edits quietly. Jinshu marks every revision word by word and applies it only after you confirm; Guanlan puts its output in a draft folder until you approve it into the library.
- We build our own foundations: Guanlan's search runs on our own data engine, libbitcask. No external database service — install and it just works.
Company
Tongjiu (Dalian) Internet Technology Co., Ltd.(通九(大连)互联科技有限公司)
Website: ttalk.im
Contact us: [email protected]