Deep Research

brahma-research is an agentic research pipeline, not a single model. Given a question, it plans an approach, searches the web, reads the most relevant pages, and synthesizes a cited report — driven by brahma-30b-reasoning with tool use. Because a thorough answer can take minutes and many tool calls, research runs as an asynchronous job: you start it, then poll for the result.

POST /api/v1/voice-engine/research        # start a job  -> 202 { id, status }
GET  /api/v1/voice-engine/research/{id}    # poll         -> status, then report + citations
GET  /api/v1/voice-engine/research         # list your org's jobs

Start a job

curl https://api.stage.peachdeskai.com/api/v1/voice-engine/research \
  -H "X-API-Key: $PEACHDESK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Compare the leading open-weight multilingual embedding models for Indian languages in 2026, with sources.",
    "max_tool_calls": 12
  }'

Returns immediately:

{ "id": "3f9c…", "status": "pending", "created_at": "2026-07-29T10:00:00Z" }

Field	Type	Notes
`query` (required)	string	The research question.
`max_tool_calls`	integer (1–25, default 12)	The cost/depth knob — the hard cap on web searches + page reads before the agent must synthesize.
`model`	string	Driver model; defaults to `brahma-30b-reasoning`.

Poll for the result

curl https://api.stage.peachdeskai.com/api/v1/voice-engine/research/3f9c… \
  -H "X-API-Key: $PEACHDESK_API_KEY"

status moves pending → running → completed (or failed). While running, tool_calls_used climbs so you can watch progress. When completed, the response includes the report and citations:

{
  "id": "3f9c…",
  "status": "completed",
  "query": "Compare the leading open-weight multilingual embedding models…",
  "report": "Three models lead for Indian-language retrieval in 2026 [1]. …",
  "citations": [
    { "url": "https://…", "title": "MTEB multilingual leaderboard", "start": 41, "end": 44 }
  ],
  "tool_calls_used": 9,
  "created_at": "2026-07-29T10:00:00Z",
  "updated_at": "2026-07-29T10:02:14Z"
}

Each citation’s start/end are character offsets of a [n] marker in report, so you can render footnotes or link the exact spans back to their source. Poll every few seconds; a typical job finishes in one to a few minutes depending on max_tool_calls.

How it works

Plan — the reasoning model breaks the question into what it needs to find.
Search — it issues web searches through your configured provider.
Read — it fetches and extracts the readable text of the most promising pages.
Synthesize — it writes a report and cites each source as [1], [2], …

Every fetched URL is screened before the request is made (only public http/https hosts — private/loopback/metadata addresses are blocked), tool arguments are validated, and the loop stops at max_tool_calls. Jobs, reports, and citations are stored per organization and metered as the underlying brahma-30b-reasoning reasoning tokens.

Prompting tips

Ask a real question, not a keyword. “What changed in X between 2024 and 2026, and why does it matter for Y?” gets a far better report than “X”.
State the deliverable. Ask for a comparison table, a decision recommendation, or a timeline — the agent will structure the report to match.
Tune depth with max_tool_calls. Raise it for broad, multi-source questions; lower it for quick fact-finding to save time and cost.

Deep Research requires a web-search provider to be configured for your deployment (self-hosted SearXNG or a Tavily key). If it isn’t configured, starting a job returns a clear error. Research jobs run in the background on your API service; a very long job is best started early and polled.

Resources

Authentication & Errors

Voice Engine API

Bibha BRAHMA LLM

Bibha VAANI TTS

Bibha LIPI STT

Start a job

Poll for the result

How it works

Prompting tips

​Start a job

​Poll for the result

​How it works

​Prompting tips

Start a job

Poll for the result

How it works

Prompting tips