# 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 (report writer, findings extractor, linker) grounded on the passages retrieved from that collection to produce a thorough, cited markdown report with key findings, a confidence score and source diversity. 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; a depth above 1 also follows the most relevant links found inside crawled pages) 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 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, and the crawl interleaves their results so every angle contributes sources. - **Crawling** fetches candidates concurrently, first with a plain HTTP client, falling back to a headless browser for JavaScript-heavy pages. A readability extractor isolates the main article content of each page (navigation, cookie banners and footers are discarded). For social sites that block bots (X, YouTube, Reddit and similar) the readable text supplied by the search engine is used directly, so those sources still contribute their real content instead of a login wall. At depth above 1 the most relevant links inside crawled pages are followed. Identical content is de-duplicated, and a cross-session URL cache tracks 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** retrieves the passages most relevant to your question from that collection and runs the report writer, findings extractor and linker agents to write the cited report, extract findings, and score overall confidence. The score combines confidence, source diversity and coverage. If synthesis fails, the report page marks the run as degraded instead of presenting raw source material as a report. ## 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](/docs/tools.html) API group for request and response shapes.