[ engine · AIQ Deep Research ]

AIQ Deep Research

Multi-step research powered by the NVIDIA AI Q&A (AIQ) Deep Research blueprint: planner → researcher → synthesizer → citer. Fans sub-queries out to Tavily + Finnhub + EDGAR in parallel, then produces a Bloomberg-grade markdown note with inline citations.

What it is · how it works · why it matters

[ what ]

Multi-step research using NVIDIA's AI Q&A (AIQ) Deep Research blueprint. Plans sub-queries, fans them out in parallel, synthesizes a Bloomberg-grade note with inline citations.

[ how ]

Pipeline: planner → researcher → synthesizer → citer. Planner (Nemotron Super) decomposes into 3-5 sub-queries. Researcher fans out to Tavily / Finnhub / EDGAR in parallel. Synthesizer joins; citer inserts inline links.

[ why ]

Single-shot LLM answers tend to produce ungrounded citations. The blueprint's planner/synthesizer separation grounds each claim in retrieval results. Our internal eval gold set measures factuality differences vs. single-shot baselines.

Overview

Two engines, picked at request time:

AIQ Deep Research adapter (literal blueprint) — straight port of the NVIDIA AI Q&A Deep Research blueprint. Uses the blueprint's planner + synthesizer prompts unchanged.
In-house variant — adds finance-specific tool calls (Finnhub fundamentals, EDGAR filings, news ranking) on top of the same Deep Research skeleton.

Both live under src/traderspace/agents/aiq_deep_research_adapter.py and deep_research_agent.py. The Research workbench picks the in-house variant by default; the chat tool deep_research can pick either.

Pipeline

question
  ↓
PLANNER (Nemotron 3 Super)
  produces N sub-queries, each tagged with a source
  ↓
RESEARCHER (parallel fan-out)
  for each sub-query:
    pick the right source (Tavily for web, Finnhub for fundamentals, EDGAR for filings)
    pull top K results
    summarize per-result
  ↓
SYNTHESIZER (Nemotron 3 Super)
  joins all summaries into one coherent markdown
  ↓
CITER
  inserts inline citation links keyed back to source URLs / filing IDs
  ↓
markdown answer + sources list

How to use it

From the Research page

Type a question into the DeepResearch box. The agent streams its plan, then each sub-query result as it arrives, then the synthesized answer.

From chat

Ask any sufficiently broad question in the embedded chat. Kimi will route to deep_research if the question looks research-shaped (multi-source, multi-step).

From the REST API

curl -X POST http://127.0.0.1:8015/api/research/deep \
     -H 'content-type: application/json' \
     -d '{"symbol":"NVDA","question":"What is driving short interest the last 5 days?"}'

What good answers look like

Replies are evaluated on citation grounding. A well-formed DeepResearch reply:

Has at least 3 inline citations.
Names the source by domain (e.g. "(seekingalpha.com)", "(NVDA 10-Q · 2025-04-30)").
Contradicts itself across sources when sources disagree — and says so explicitly.
Includes a "what would change my mind" closing paragraph.

If the answer is one long paragraph with no inline citations, the synthesizer pass lost the citation map — usually because the researcher pass returned too many short results for the synthesizer prompt window. Rerun with a tighter question.

Tuning the pipeline

Planner sub-query count — N_SUBQUERIES in the agent file. Default 4. Raise for harder questions.
Per-source results — K_RESULTS. Default 3 Tavily, 5 Finnhub news.
Model — defaults to Nemotron Super 120B; switch via the model arg.

REST surface

Verb	Path	Purpose
POST	`/api/research/deep`	Body: `{symbol, question, engine?}`. Streams SSE.
GET	`/api/research/deep/history?limit=10`	Recent runs (audit).