Tutorial
How to chat with your PDFs and get cited answers
The answer is in one of the PDFs. Probably. In Conduit, a free, open-source desktop AI app, you put the documents in a collection once, then ask questions across all of them and see which document and which passage every answer came from.
Introduction
Attaching a PDF to a chat works for one document. For a set of documents you keep coming back to, such as policies, contracts, manuals or research papers, attaching them all to every chat is slow and soon runs into the model's limits, and the answer doesn't say where it came from.
Conduit's Documents solve this with collections. When you add a file to a collection, Conduit reads it and splits it into sections of about 1,000 characters, and an embedding model indexes each section so it can be found by meaning as well as by its words. When a chat has a collection attached, each question searches it, the six best-matching sections go to the model with your message, and the documents they came from are shown under your question. Click one to read the passage.
In this tutorial, you will put three PDF policies of a company called Northwind in a collection: an
employee handbook, a travel and expense policy and a benefits guide. You will ask one question that
needs two of them, check the cited passages, and point a follow-up at a single document with a #
reference.
Key takeaways
- A collection is a group of files, indexed once, that any chat can search. Add files by dragging them onto the window.
- Attach a collection to a chat with + and then Documents…. Each question then searches it.
- The documents and sections used for an answer appear under your question. Click one to read the passage.
- Each question gets the six best-matching sections. For a question about one document, type
#and name it. - The text of your documents goes to the provider that indexes the collection, unless you index it with Ollama on your own computer.
Prerequisites
To follow this tutorial, you will need:
- Conduit installed on Windows, macOS or Linux. Download the latest release from the releases page.
- A provider that can index documents: OpenAI, Gemini, OpenRouter or Ollama. Anthropic doesn't offer one. This tutorial indexes with OpenRouter, which uses OpenAI's
text-embedding-3-small, and answers with GLM 5.3 Flash. To set a provider up, see the provider guides. - A few PDFs with a text layer: ones you can select text in. Scanned PDFs can't be added.
Step 1 — Creating a collection
A collection is indexed by the provider that is active when you create it, and keeps that provider. In this step, you will create one.
First make the indexing provider the active one: pick one of its models in the composer of any chat. Here that is any OpenRouter model.
Then choose Context in the side rail, and Documents at the top. The page lists your collections
on the left. Choose New collection, enter a name, here Northwind policies, and confirm. The new,
empty collection opens, with Embedded by OpenRouter · openai/text-embedding-3-small under its name.
Next, you will add the PDFs.
Step 2 — Adding the PDFs
In this step, you will add the three PDFs and let Conduit index them.
Drag the files from File Explorer or Finder anywhere onto the Conduit window, or choose Import document to pick one file at a time. A dropped batch asks where it should go:
Check that Collection names the right one and choose Add. The first time you add a PDF, Conduit explains what to expect:
The PDF is read on your computer. What can go wrong is the order of the text: a two-column page or a table can come out jumbled, which you will see in the passages later. Choose Continue.
The first time you add documents to a collection indexed by a cloud provider, Conduit also asks Send document text to OpenRouter? Indexing sends the text of every document you add to that provider, so agree only if that's acceptable for these documents. The page then says Document text is sent to OpenRouter for indexing, with a Revoke button to withdraw your agreement.
Each file shows Reading the document…, then Indexing… with a count of sections, and is then listed with its size, type and number of sections:
Next, you will attach the collection to a chat.
Step 3 — Attaching the collection to a chat
A chat only searches the collections you attach to it. In this step, you will attach one.
Choose Chats and New chat. In the composer, choose + and then Documents…. The panel lists your collections, each with a switch. Turn on Northwind policies:
The arrow next to a collection opens its list of documents, where you can leave single documents out of this chat. Choose Close. The collection now shows as a chip above the message box, and every question in this chat searches it until you remove the chip.
Next, you will ask a question.
Step 4 — Asking a question across the documents
In this step, you will ask a question whose answer is spread over two documents.
Send:
I'm flying to Tokyo next month for a client workshop, and the flight is 11 hours.
Which class can I book, what's the hotel limit, and how much can I spend on meals
each day? Also, I have 7 vacation days left this year. How many can I carry over,
and by when do I have to use them?
The flight, hotel and meal rules are in the travel policy; the vacation rules are in the handbook. The answer covers both:
The chips under your question are the sources: each document the answer drew on, with the collection and the sections used. Every figure in the answer is in those sections: Business class for flights over 9 hours, $320 a night in Tokyo, $75 a day for meals, and up to 5 days carried over until 31 March.
Next, you will check one of them.
Step 5 — Checking the cited passage
An answer is only as good as the passage behind it. In this step, you will read it.
Click the handbook chip. The passage opens, with the section number, the other sections that were used, and the file's location on your computer:
Section 3 holds the carry-over rule word for word. This is also where you see how the PDF was read: the section starts partway through the vacation table, because sections overlap a little so a sentence split between two of them is still found.
When an answer matters, open its source. If the passage doesn't say what the answer says, the answer is wrong, whatever the model's confidence.
Next, you will ask about a document the search didn't reach.
Step 6 — Pointing a question at one document
Each question gets the six best-matching sections across the collection. A question with many parts, or one worded differently from the document, can miss a document that has the answer. In this step, you will name the document instead.
In the same chat, type # and the start of the document's name. A list of matching documents opens:
Pick northwind-benefits-guide-2026.pdf. It becomes a chip above the message box, and the search for this message is limited to that document. Then type the question and send it:
Am I covered if I get sick on the Tokyo trip, and do I need to buy travel insurance?
The answer comes from the benefits guide: business travel insurance, paid by Northwind, covers
medical emergencies on approved trips. A # reference works with any document in any collection, even
one this chat doesn't have attached.
Which to use: attachment, collection or folder
Conduit has three ways to give a chat your files. They suit different jobs:
| Attach a file | Document collection | Workspace folder | |
|---|---|---|---|
| Best for | One document, in one chat | A set of documents you ask about again and again | Files the assistant should read and work with |
| How the model sees it | The whole text, or the PDF itself for models that read PDFs | The six best-matching sections per question | Files it chooses to open, a page at a time |
| Shows sources | No | Yes, document and section | No |
| How to add | + → Attach files…, or drag onto the chat | Context → Documents | + → Workspace folder… |
To turn a single document into slides instead, see How to turn a PDF into a slide deck with AI.
Document limits at a glance
| Limit | |
|---|---|
| File types | PDF, Word (.docx), Markdown, plain text, CSV, HTML |
| File size | Up to 32 MB |
| Reading time | A file that takes longer than 30 seconds to read is skipped |
| Scanned PDFs | Not added: there is no text to read |
| Section size | About 1,000 characters, overlapping by 150 |
| Sections per question | The six best matches |
| Indexing providers | OpenAI, Gemini, OpenRouter, Ollama |
Questions
Can I keep my documents on my own computer?
Yes, with Ollama. A collection is indexed by the provider that was active when you created it, and only OpenAI, Gemini, OpenRouter and Ollama can index documents. Make Ollama the active provider, with the nomic-embed-text model pulled, before you choose New collection, and the document text is indexed on your computer. Answer with a local model too, and nothing leaves it. With a cloud provider, the text of every document you add is sent to that provider to be indexed, and Conduit asks you once before the first import.
Do the citations give page numbers?
No. They give the document and the section: a passage of about 1,000 characters, numbered in order through the document. Click a citation to read the passage itself. PDFs are turned into text when they are added, so page numbers are not kept.
Can I add scanned PDFs?
No. A scanned PDF has no text layer, so nothing can be read from it, and Conduit says so and doesn't add it. Run it through OCR first, or attach it to a single chat with a model that reads PDFs directly, such as Claude, GPT-5 or Gemini.
What does it cost?
Conduit is free. Indexing is billed by the provider that indexes the collection; three short PDFs cost a fraction of a cent with OpenAI's text-embedding-3-small through OpenRouter. Each answer costs what any chat message costs: the two answers in this tutorial cost $0.0004 together on GLM 5.3 Flash.
Which other file types can I add?
Word documents (.docx), Markdown, plain text, CSV and HTML files, as well as PDFs, up to 32 MB each.
Conclusion
In this tutorial, you put three PDFs in a document collection, attached it to a chat, asked a question
that needed two of them, read the passage behind the answer, and pointed a follow-up at a single
document with a # reference.
The same steps work for any set of documents you keep asking about: contracts, product manuals, research papers, meeting notes, a codebase's design documents. To go further, explore these guides:
- How to turn a PDF into a slide deck with AI, to present what a document says.
- How to run AI agents side by side in Conduit, to give an agent a collection it searches on every task.
- The provider guides, to choose the models that index and answer.
Conduit details match the latest release and are checkable in the documentation and the source. Found an error? Tell us and we will fix it.
Related
- How to turn a PDF into a slide deck with AI — Turn a PDF report into an editable slide deck with AI: attach the PDF, approve a storyline, build the slides, edit the words and export to HTML, PowerPoint or PDF.
- How to set up MCP servers on Windows — Set up MCP servers on Windows: install Node.js and uv, add local servers with npx or uvx, add remote ones by URL or from the registry, fix common errors.
- How to run AI agents side by side in Conduit — Define AI agents once, each with its own model and tools, then hand them work from chat, watch them work side by side, have their work reviewed and run them on a schedule.
- How to do deep research with Conduit and Exa — 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.
- How to automate tasks with Conduit Workflows — 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.