> Markdown twin of https://conduitllm.com/research-with-exa.html > Turn on Exa search, write a research brief and read a cited report in Conduit, with solar cell research as the worked example. No Exa key needed. --- Tutorial # How to do deep research with Conduit and Exa Updated 8 October 2026 A good research question rarely has a one-line answer. Answering it means searching, reading dozens of pages, keeping track of who said what, and noticing where sources disagree. Research in [Conduit](https://conduitllm.com/), a free, open-source desktop AI app, does that work for you and hands back a report in which every claim cites a source, and every citation has been checked against the words on that page. ## Introduction In this tutorial, you will use Conduit's Research feature with [Exa](https://exa.ai) as its search engine to answer a real engineering question: *how close are perovskite-silicon tandem solar cells to commercial use?* You will turn on Exa search without creating an Exa account, review and edit the research plan before anything runs, follow the run as it reads pages, and finish with a cited report you can check line by line. The solar question is a good test because the details matter. Efficiency records, lifetime claims and factory announcements come from very different kinds of sources — certified lab measurements, peer-reviewed papers, company press releases — and a useful report has to keep them apart. ## Key takeaways - Conduit uses Exa as its default search engine through [Exa's hosted MCP server](https://exa.ai/docs/reference/exa-mcp), which Exa offers free without an API key. - Research starts with a brief — sub-questions, scope and depth — that you edit and approve before any search runs. - Every claim in the report carries a quote from its source, and Conduit's own code, not the model, confirms the quote is on the page before the claim is used. - The finished report is a Markdown document with findings per sub-question, open questions and numbered sources. - You pay only for the model you choose; Conduit is free, and keyless Exa search costs nothing. ## Prerequisites To follow this tutorial, you will need: - **Conduit installed** on Windows, macOS or Linux. Download it from the [releases page](https://github.com/runningpixels/conduit/releases). - **A model provider set up in Conduit** with your own API key. Research uses the model you have selected. To set one up, follow one of the [provider guides](https://conduitllm.com/providers.html), such as [Use OpenRouter in a desktop app](https://conduitllm.com/providers/openrouter.html). - **Local-only mode turned off.** Research searches and reads the web, so it is not available in local-only mode. You can check this in **Settings → Privacy & data**. - **(Optional) An Exa API key.** You don't need one for this tutorial. If you plan to research often, a key from the [Exa dashboard](https://dashboard.exa.ai/api-keys) raises Exa's rate limits. ## Step 1 — Turning on web search with Exa Research finds its sources through Conduit's local search backend, so the first thing you will do is turn on web search and confirm that Exa is the backend in use. [Exa](https://exa.ai) suits this job because Research doesn't want a search engine's answer; it wants pages to read. Exa is a search API built for AI applications: it describes its service as a web-scale index with its own retrieval models, and each result comes with highlights, the passages Exa judges most relevant to the query. Conduit asks Exa for ten results per search and uses those highlights to decide which pages to read in full. Open **Settings → Web search** and tick **Enable web search**. The first time you do this, Conduit shows a notice titled **Enable web search?** that explains where your search queries go. Read it, then choose **Allow**. Next, look under **Local search backend** and make sure **Exa** is selected. Exa is the default, so on a fresh install it already is. Leave the **Exa API key (optional)** field empty. **Note:** Search queries go from your machine to Exa, and Exa may keep them under its own terms. [Exa's privacy policy](https://exa.ai/privacy-policy) says it uses query data to improve its products, including training models, so leave confidential details out of your research questions. Web search is now on, with Exa as the backend. In the next step, you will ask your research question and start a run. ## Step 2 — Starting a research run A Research run begins like an ordinary chat message, with Research switched on for that message. In this step, you will turn it on and ask the solar question. Open a new chat. In the message box, open the **+** menu and choose **Research**. A **Research** chip appears in the message box, which means your next message starts a Research run instead of a normal reply. Type the following question and send it: How close are perovskite-silicon tandem solar cells to commercial use? **Note:** You can also start from Home. Type the question in the ask box and choose the **Research** chip; Conduit opens a new chat and starts the run there. If the **Research** item is greyed out, hover over it to see why — usually web search is off or local-only mode is on. Conduit doesn't search yet. Instead, it drafts a plan for the run, which you will review in the next step. ## Step 3 — Editing the research brief Before anything is searched, Conduit shows a **Research brief**: the sub-questions it will answer, an optional scope, and a depth. The sub-questions decide what gets searched, so this is where a few seconds of editing have the most effect on the report. For the solar question, a useful brief covers these four sub-questions: - What is the record efficiency for perovskite-silicon tandem cells, and who holds it? - How long do they last outdoors, and what limits their lifetime? - Which companies are manufacturing them, and at what scale? - What testing and certification do they need before they are sold? Edit the drafted sub-questions until they match your intent. Use **Add a sub-question** to add one and **Remove** to drop one; a brief holds between one and six. The second sub-question matters most here: efficiency is the headline number, but lifetime is what decides whether a solar panel can be sold. In **Scope (optional)**, enter `2024–2026`. This keeps the run focused on current results rather than older records that have since been broken. Finally, choose a depth. Each one sets a ceiling on how far the run may go: | Depth | Searches | Pages read | Time | | --- | --- | --- | --- | | Quick | up to 8 | up to 15 | about 10 min | | Standard | up to 20 | up to 40 | about 20 min | | Deep | up to 50 | up to 100 | about 40 min | Choose **Standard**. Four sub-questions with forty pages to read between them is enough for this topic, and the run finishes in under half an hour. Then choose **Start research**. You have approved the brief, and the run has started. In the next step, you will follow its progress in the chat. ## Step 4 — Following the run While Research works, a card in the chat shows what it is doing. In this step, you will read that card and learn what happens at each stage. The card's status line moves through these stages, repeating some of them as needed: - **Searching the web…** — Conduit sends search queries to Exa for each sub-question. - **Reading pages…** — it reads the most promising results in full, three at a time, including PDFs up to 20 MB. The card shows the site it is reading. - **Extracting claims…** — a model reads each page and pulls out claims, each with the exact words from the page that support it. - **Checking for gaps…** — if a sub-question is still unanswered, Research searches again. - **Writing the report…** and **Checking quotes against sources…** — the report is written from the checked claims and reviewed. Under the status line, the counters for **Searches** and **Pages read** show how much of the budget is used, for example `7 of 20`, and **Claims found** shows how many claims have been extracted so far. The run shown in this tutorial used GLM 5.3 Flash, a low-cost model, through OpenRouter, and finished in a few minutes. For the solar question, much of the best evidence sits in national laboratory reports and research papers, many of them PDFs. Because Research reads whole documents rather than search snippets, it can use those sources directly. **Note:** You can choose **Stop** at any point. Research doesn't throw its work away: it writes a report from the claims it has verified so far, with a note at the top saying the run was stopped early. Once the card shows that the run is done, the report is ready. Next, you will open it and read the findings. ## Step 5 — Reading the report When the run finishes, the card shows a short summary, how many sources were cited and read — for example `8 cited · 8 read` — and an **Open report** button. In this step, you will open the report and see how it is organised. Choose **Open report**. The report opens beside the chat, with these sections: - **Summary** — the short answer to your question. - **Findings** — one part per sub-question, with numbered citations such as `[3]` after each statement. - **Where sources disagree** — included when sources give conflicting answers. - **Open questions** — sub-questions the run couldn't answer, naming any sites it couldn't read. - **Sources** — the numbered list of cited pages, each with the date it was fetched, followed by pages it read but didn't cite. If the card says that some claims were left out because their quotes could not be found in their sources, that is the quote check doing its job, which you will look at in the next step. ## Step 6 — Checking a claim against its source A cited report is only useful if the citations hold up. In this step, you will check one claim yourself and see how Conduit has already checked all of them. Pick a statement in the **Findings** section — the record efficiency is a good choice — and note its citation number. In the card, expand the source list and choose that source. The page opens in your browser, where you can find the passage the claim is based on. Conduit has already run the same check on every claim before writing the report: 1. **Every claim carries a quote.** The model that reads a page must return the exact words that support each claim. 2. **Code looks for those words on the page.** Conduit's own code, not the model, searches the stored text of the page for the quote — at least 20 characters of it, ignoring case, spacing and curly quotes. A claim whose quote isn't there is dropped. 3. **Only checked claims reach the writer.** The citation numbers and the source list are produced by the app, not typed by the model. 4. **A review pass reads the draft.** It looks for sentences that say more than their source, or state a fact without a citation, and corrects them. 5. **Weak sources are labelled.** Each page gets a credibility rating, and a low-rated source is marked in the source list with the reason. Official sites, such as government laboratories, are never rated low. For a topic like solar cells, where a press release and a certified laboratory measurement can give very different numbers, these checks are what make the report something you can rely on and verify. ## Questions **Do I need an Exa API key?** No. Conduit reaches Exa through Exa's hosted MCP server, which Exa offers free without a key, with rate limits. If a run reports that Exa's free search is busy, wait a minute or add a key in **Settings → Web search**; Exa says a key raises the limits, and its free tier gives new accounts $10 of credit, topped back up to $10 each month. **What does a research run cost?** Conduit is free, and keyless Exa search is free. What you pay for is the model: every planning, reading and writing step is a call to the provider you picked, billed to your own key. Each depth has a token ceiling — 150,000 tokens for Quick, 400,000 for Standard and 1,000,000 for Deep — and the run stops and writes up what it has when it reaches it. Every call is recorded in Usage & Cost. **Does Research use my model's own web search?** No. Research always searches through the local search backend you choose in Settings, even when the search source for ordinary chats is set to the provider's hosted search. Hosted search answers questions; Research needs a list of pages it can read and quote. **Can I use Tavily or Brave instead of Exa?** Yes. Pick Tavily, Brave Search or your own SearXNG under **Local search backend**. Tavily and Brave need an API key; SearXNG needs your instance's address. DuckDuckGo is also listed, but it only returns encyclopedia-style snippets, so it is a poor fit for Research. **Can a web page take over the research run?** No. The text of a page only ever reaches a model call that has no tools, so a page that contains instructions cannot make Conduit search, fetch or save anything. Claims whose text looks like an instruction are dropped, and Conduit only fetches public https addresses that came from the search results. ## Conclusion In this tutorial, you turned on web search with Exa as the backend, asked a research question about perovskite-silicon tandem solar cells, shaped the run with a brief, followed it as it read pages, and checked the cited report against its sources — all without an Exa account. From here, try Research on a question from your own field, or choose **Write from this report** on the card to turn the report into a draft article that cites the same verified facts. To go further, explore these guides: - [The Research documentation](https://conduitllm.com/docs.html#research), for every setting and limit. - [Chat with your documents](https://conduitllm.com/use-cases/chat-with-documents.html), to get cited answers from your own files instead of the web. - [The provider guides](https://conduitllm.com/providers.html), to choose and set up the model that runs your research. [Download Conduit](https://github.com/runningpixels/conduit/releases) Sources for the Exa facts, checked 6 October 2026: Exa's [home page](https://exa.ai) (index and highlights), [MCP server documentation](https://exa.ai/docs/reference/exa-mcp) (keyless use, rate limits, API keys), [pricing](https://exa.ai/pricing) (free tier) and [privacy policy](https://exa.ai/privacy-policy) (query data). Conduit details match the latest release and are checkable in [the documentation](https://conduitllm.com/docs.html) and [the source](https://github.com/runningpixels/conduit). Found an error? [Tell us](https://github.com/runningpixels/conduit/discussions) and we will fix it. ## Related - [How to automate tasks with Conduit Workflows](https://conduitllm.com/automate-with-workflows.html) — Build a Conduit workflow that watches a project's releases, summarizes each with a low-cost model, saves the summary and alerts you only to breaking changes. - [API pricing vs a monthly subscription](https://conduitllm.com/api-vs-subscription.html) — How many messages a day before pay-per-token beats a $20 subscription? A calculator, plus why conversation length moves the answer more than volume does. - [How to demo an MCP server to a customer](https://conduitllm.com/mcp-distribution.html) — Hosting an MCP server is not a customer demo. Compare a hosted chat, Claude Desktop with an extension, and your own client — and what you can ship today.