DeepSearch

DeepSearch is a multi-agent deep web researcher. Given a single research question it plans a set of
diverse web search queries, crawls and reads the most relevant sources, indexes everything into a
private vector collection for that run, then runs a chain of agents (summarizer, critic, linker) to
produce a grounded, cited report with a confidence score, source diversity and an explicit list of
open gaps. Progress streams live while it works, and afterwards you can chat with the gathered
evidence.

Open it from the Tools menu, or go straight to /tools/deepsearch. It is public: you do not
need an account.

Running a research job

  1. Type a focused research question.
  2. Set the depth (1-4) and the maximum number of pages to crawl (up to 30).
  3. Press Research. Progress appears immediately: query planning, web search, crawling each
    source, indexing, then the analysis agents.
  4. You can pause, resume or cancel a run at any time.
  5. When it finishes, open the report to read the summary, findings, gaps and sources, and to chat
    with the research.

You can run one job at a time. Targets that resolve to private or local addresses are refused, and
every fetched URL (including redirects) is checked.

How it works

  • Query planning expands your question into several complementary searches.
  • Crawling fetches each candidate first with a plain HTTP client, falling back to a headless
    browser for JavaScript-heavy pages. Identical content is de-duplicated, and a cross-session URL
    cache avoids re-fetching pages seen by earlier runs.
  • Indexing splits each page into overlapping chunks, embeds them through the AI gateway (with a
    local embedding fallback when the gateway is unavailable), and stores them in a per-session
    ChromaDB collection.
  • Analysis runs the summarizer, critic and linker agents to synthesise findings, surface gaps,
    and score overall confidence. The score combines confidence, source diversity and coverage.

Chatting with the research

Every finished session has a chat pane. Answers are grounded only in the sources captured during
that run, using hybrid retrieval (vector similarity plus keyword/BM25 ranking) over the session
collection, and every claim is cited back to a source.

Exporting

A finished report can be downloaded as Markdown, JSON or PDF from the report page.

Ask Devii

You can also run research in plain language through the Devii assistant:

Run a deep search on the history of the transistor and summarise the findings.

Devii queues the job, polls it, and reports the score, confidence and a link to the report.

Programmatic access

The same research is available over the API: POST /tools/deepsearch/run to queue,
GET /tools/deepsearch/{uid} to poll, and GET /tools/deepsearch/{uid}/session for the full report
(HTML or JSON). See the Tools API group for request and response shapes.

Open DeepSearch