Deepresearch Search

Deep Research Search (Stage 2b)

Takes the Stage-0 brief and uses the Gemini Deep Research Agent to autonomously plan, search, read, and synthesize the prior literature, then parses the resulting cited report into <stem>.json + <stem>.bib for dedup and screening. It complements the other deep searches: Undermind and Scholar Labs are browser-driven; this one is a pure API call.

The stage is two halves:

• scripts/deepresearch_search.py — calls the Interactions API (deep-research-max-preview-04-2026 by default), runs the task in the background, polls to completion, and saves the report + raw response.
• scripts/deepresearch_ingest.py — UI-independent parsing. It reads the agent's KEY PAPERS section (and, as a fallback, citations in the raw response), best-effort-enriches via Crossref, and writes the pipeline JSON. Importable, and runnable standalone on a saved report.

The flow

brief → wrapped into a literature-review prompt that ends with a parseable KEY PAPERS section (Title | Authors | Year | Venue | DOI or URL, one per line) → POST /v1beta/interactions with background=true, store=true, agent deep-research-max-preview-04-2026 → poll GET /v1beta/interactions/{id} until completed → save deepresearch_report.md + deepresearch_raw.json → parse the KEY PAPERS lines → enrich → stage2b_deepresearch.json (+ .bib).

Deep Research is asynchronous and takes minutes (the API caps a task at 60 minutes; most finish under 20). Google lists "literature reviews" as a primary use case for the agent.

Why an API, not a browser

When this pathway was first built, "Google deep research" had no usable API and was retired in favour of Scholar Labs. Google has since shipped the Deep Research Agent on the Interactions API (deep-research-preview-04-2026 and deep-research-max-preview-04-2026, on Gemini 3.1 Pro), so the stage is now a clean, scriptable API call. It is a retrieval service (like SearchAPI), not part of the pipeline's own reasoning, so it is compatible with the agent-driven run's no-Anthropic-API rule.

API key

Stored in ~/.lit-review-pipeline.env (gitignored). If it is missing the stage degrades gracefully.

Cost

Pay-as-you-go per task (it is an agentic loop, not one request):

• Deep Research (deep-research-preview-04-2026): ~$1–3 per task.
• Deep Research Max (deep-research-max-preview-04-2026, the default here): ~$3–7 per task — more searches and synthesis, better for literature coverage.

CLI

# Driven by the orchestrator (the normal path)
python scripts/deepresearch_search.py --query-file undermind_brief.txt \
    -o stage2b_deepresearch.json --debug-dir debug_deepresearch

# Re-parse a saved report (no API call)
python scripts/deepresearch_ingest.py --report debug_deepresearch/deepresearch_report.md \
    --raw debug_deepresearch/deepresearch_raw.json -o stage2b_deepresearch.json

Graceful degradation

If GEMINI_API_KEY is missing, the task fails or times out, or no papers can be parsed, the driver prints DEEPRESEARCH_DEFERRED, writes an empty result list, and exits 0. The orchestrator marks the stage deferred and the rest of the pipeline still completes.

Source extraction

The reliable channel is the KEY PAPERS section the prompt asks the agent to emit, because the Interactions API does not support structured outputs. Each line is Title | Authors | Year | Venue | DOI or URL. The ingest also scans the raw response for (title, url) citations as a fallback when the section is thin. Records carry source: "deepresearch"; missing DOIs/abstracts/journals are filled at the dedup stage's enrichment cascade (OpenAlex → Crossref → S2).

Output schema

Same as the other stages, with source: "deepresearch":

{"title": "...", "authors": "A, B", "year": "2024",
 "doi": "https://doi.org/10.x/y", "abstract": "", "journal": "...",
 "url": "...", "source": "deepresearch", "verified": false,
 "citations": 0, "open_access": false}

Notes / limitations

• The Interactions API is in public beta; response schemas may change. The driver dumps deepresearch_raw.json so the ingest (and you) can adapt if the citation structure shifts; the KEY PAPERS text channel is schema-independent.
• background=true is required and implies store=true; both are set.
• Default tools are Google Search + URL Context + Code Execution.