MCP endpoint: https://mcp.insap.ai/mcp (Streamable HTTP). Tools, resources and prompts can be listed without a key; running a tool needs an API key from the VoiceAxis team (Authorization: Bearer <key>).
# Gyani: user guide
Gyani is VoiceAxis's conversational language model. It is fast, Hindi-first and English-capable, and it is
built for the short, natural replies a voice agent or a support desk needs.
## Use it for
- Customer-care replies in **Hindi (Devanagari), English or Hinglish**, using facts you supply
- Summarising calls, chats and notes; pulling out the intent, next action, names, numbers and dates
- Rewriting text: politer, shorter or simpler, or switched between Hindi and English
- Q&A over a passage you paste in
## Don't use it for
- Long multi-step reasoning, maths, code, or legal or medical advice
- Facts it has not been given. It does not browse the web, and like any small model it can be wrong about the
world. Put the facts in the request.
- Tool or function calling (not enabled)
## Tools
| Tool | Needs a key | What it does |
|---|---|---|
| `gyani_status` | yes | Is Gyani online right now? Also returns its limits. Does not run the model. |
| `gyani_chat` | yes | Send `prompt` (plus optional `system`, `history`, `max_tokens`, `temperature`) and get `reply` back |
Listing tools, resources and prompts, and reading these docs, needs **no key**. Running a tool needs
`Authorization: Bearer <key>`. Keys are issued per person or partner by the VoiceAxis team and can be revoked.
## Limits
- **Availability:** Gyani runs on a GPU server that is switched off when idle. When it is off, tools return
"Gyani is offline right now". Check with `gyani_status`.
- **Size:** up to 24,000 characters in (about 8,000 tokens) and up to 2,048 tokens out per call.
- **Rate:** 30 tool calls per minute per key. Gyani handles a few conversations at a time, shared with live
voice calls, so busy moments can add a short wait.
- **Speed:** the first token arrives in about 0.1 s, then roughly 40-60 tokens per second.
## Tips
- Say the language you want in `system`: "Reply in Hindi (Devanagari)", "Reply in English", or "Reply in
Hinglish as the customer writes".
- Ask for the length you want ("one or two sentences"). Gyani follows length instructions well.
- Use `temperature` 0-0.3 for factual work and 0.7 or above for creative wording.
## Example
`gyani_chat` with
`system = "You are a polite support agent. Service visits happen within 48 hours. Reply in Hindi, one sentence."`
and `prompt = "मेरा गीज़र खराब हो गया है, कोई कब आएगा?"`
returns a reply like: "जी, हमारे टेक्नीशियन 48 घंटे के अंदर आपके पास पहुँच जाएँगे।"
## Prompts (templates)
- `hindi_customer_reply(customer_message, business_facts)`
- `summarise_call(transcript, language)`
- `rewrite_polite(text, language)`