@gagandeep023/support-chat-cli

Command line tools for support-chat: run the whole thing locally, check a model can do the job, and score it against your own documentation.

npx @gagandeep023/support-chat-cli help

dev, with nothing installed

npx support-chat dev

No database, no Redis, and no API key. It ingests a sample document, serves a page with the widget on it, and answers questions. With no model configured the replies are scripted, so you can watch the entire loop, ask something the docs do not cover, and queue for a human before spending anything.

support-chat dev
  Open       http://localhost:4000
  Model      a scripted stand-in (no model configured)
  Embeddings none (keyword search only)
  Knowledge  1 document, 3 chunks
  Storage    in memory

Point it at real docs and a real model:

export ANTHROPIC_API_KEY=...
npx support-chat dev --docs ./docs --db ./local.db

| Flag | Default | | --- | --- | | --port, -p | 4000 | | --docs, -d | a built-in sample document | | --model, -m | from the environment | | --db | in memory | | --embeddings, -e | none, or hashed / local |

doctor, before you commit to a model

support-chat doctor --model <id> [--top-k 8]

Checks the configured model can actually do the job: streaming, a parseable escalation verdict, tool calling, and context budget. "Use any model you like" is not a feature if you have no way to find out that one of them cannot return a verdict you can parse.

eval, against your own docs

support-chat eval --cases ./cases.jsonl [--docs ./docs] [--model <id>]

{"question": "what is E4021?", "grounded": true, "escalate": false}
{"question": "what is the capital of France?", "grounded": false}
{"question": "I want to talk to a human", "escalate": true}

Scores grounding, refusal and escalation accuracy. The second case matters as much as the first: a model that answers questions your documentation does not cover is the failure mode that costs you trust.

Environment

| Variable | For | | --- | --- | | ANTHROPIC_API_KEY | the Anthropic API | | SUPPORT_CHAT_BASE_URL | any OpenAI-compatible endpoint | | SUPPORT_CHAT_MODEL | model id at that endpoint | | SUPPORT_CHAT_API_KEY | key for that endpoint, when it needs one |

That covers Moonshot (Kimi), Groq, Together, Fireworks, OpenRouter, vLLM and Ollama. With none of them set, replies are scripted so the loop still runs.

The rest of support-chat

| Package | What it is | | --- | --- | | @gagandeep023/support-chat-core | Types, wire protocol, reconnect backoff | | @gagandeep023/support-chat-server | The engine: socket gateway, retrieval, routing, host tools, stores | | @gagandeep023/support-chat-widget | Customer-facing chat: web component, React, or headless | | @gagandeep023/support-chat-agent-console | Agent side, as a React component for your dashboard | | @gagandeep023/support-chat-cli | dev, doctor, eval |

The repository README is the full walkthrough, and DESIGN.md records what was rejected and why.

Requests and feedback

Ideas and questions go to Discussions, bugs to Issues.

License

MIT