--- title: "Which Documents Can an AI Knowledge Base Read?" url: https://www.insulin.dev/blog/what-documents-an-ai-knowledge-base-can-read/ canonical: https://www.insulin.dev/blog/what-documents-an-ai-knowledge-base-can-read/ type: Blog description: "File formats, size limits, crawling, and connectors — what an Insulin knowledge base accepts, and which source is the right one to start with." --- # Which Documents Can an AI Knowledge Base Read? > Canonical HTML version: https://www.insulin.dev/blog/what-documents-an-ai-knowledge-base-can-read/ 1. [Home](/) 2. / 3. [Blog](/blog/) 4. / 5. Which Documents Can an AI Knowledge Base Read? # Which Documents Can an AI Knowledge Base Read? File formats, size limits, crawling, and connectors — what an Insulin knowledge base accepts, and which source is the right one to start with. ![Shirley Guo](/authors/shirley-guo.jpg) Shirley Guo Aug 4, 2026 ![Which Documents Can an AI Knowledge Base Read?](/images/blog/what-documents-an-ai-knowledge-base-can-read/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fwhat-documents-an-ai-knowledge-base-can-read%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Integrations. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fwhat-documents-an-ai-knowledge-base-can-read%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Integrations. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fwhat-documents-an-ai-knowledge-base-can-read%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Integrations. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fwhat-documents-an-ai-knowledge-base-can-read%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Integrations. "Summarize with Perplexity") Table of Contents - [What the format list covers](#what-the-format-list-covers) - [Beyond individual files](#beyond-individual-files) - [What to start with](#what-to-start-with) - [Indexing, and what happens after](#indexing-and-what-happens-after) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _An Insulin knowledge base accepts PDF and DOCX files and text formats including Markdown, plain text, CSV, TSV, logs, JSON, HTML, XML and YAML, up to 25 MB per file. Documents arrive three ways: uploaded or pasted, crawled from a website, or synced through a connector._ * * * “Can it read our documents?” is usually the first question, and it is the wrong one to spend long on — the format list covers most of what a business actually has. The question worth spending time on is which of your sources should go in _first_, because that decision determines whether the first answers are any good. Both are below. Formats first, since that is what people came for. * * * ## **What the format list covers** Category Formats Typical source **Documents** PDF, DOCX Policies, contracts, signed agreements, decks exported to PDF **Text and markup** Markdown, plain text, HTML, XML Wikis, exported help centres, internal docs sites **Structured data** CSV, TSV, JSON, YAML Reference tables, exported records, configuration **Logs** `.log` Diagnostic output a support agent should be able to read The ceiling is **25 MB per file**. In practice that binds on scanned PDFs and rarely on anything else — a policy document is a fraction of it, while a 300-page scanned contract set can exceed it and needs splitting. One caveat worth knowing before you load a directory: a PDF’s usefulness depends on it containing text. A scan that has never been through OCR is an image as far as any retrieval system is concerned, and it will index as an empty document rather than fail loudly. If a source is legacy paperwork, check that you can select text in it. Structured formats behave differently from prose and it is worth setting expectations. A CSV of reference data is genuinely useful — retrieval finds the row and the agent reads it. A CSV as a _database substitute_ is not; asking an agent to aggregate across ten thousand rows of a retrieved file is the wrong tool for the job. Reference tables yes, analytics no. * * * ## **Beyond individual files** Three ways to bring in material that lives somewhere else: **Website crawling.** Point it at a public site and it ingests the pages. The obvious use is your own published documentation or help centre — material that is already written, already maintained, and already the canonical answer for customer-facing questions. **Connectors.** GitHub, ClickUp, and Google Drive connect as sources, with sync status visible per connector and per document. Connectors are the right choice for anything that changes, because a connected source stays current where an uploaded copy silently ages. That difference is the one to internalise. An uploaded file is a snapshot; it will still be there, still be cited, and still be authoritative-looking six months after somebody revised the original. A connected source updates. Where a document has a living home in a system Insulin connects to, connect it rather than upload it. * * * ## **What to start with** Ignore volume. The first knowledge base should be small, current, and about one domain — the thing you can grade after a week. The best first source has three properties: 1. **Somebody already maintains it.** Ownership is what stops a base rotting. 2. **It answers questions people actually ask.** Not the most impressive document — the most asked-about one. 3. **You can tell whether an answer from it is right.** Otherwise you cannot evaluate anything. That usually points at internal runbooks and approval policy rather than the contract archive. The archive is bigger and feels more valuable; it is also mostly historical, hard to grade, and full of superseded versions that will be retrieved and cited with total confidence. Once the first base works, the material that follows tends to be the same shape: procedures somebody would otherwise explain out loud, policy with a decision in it, templates and their constraints. What to leave out is covered in [how to ground an agent](/blog/how-to-ground-an-ai-agent-in-your-documents/), and it matters as much as what to include. * * * ## **Indexing, and what happens after** Documents show their state as they are processed — indexing, then indexed — so you can see when a base is ready rather than guessing why an answer looks thin. Two habits worth forming early: **Re-check after a bulk load.** A file that failed to extract does not announce itself in an answer; it just quietly is not there. Confirm the count of indexed documents matches what you loaded. **Treat the base as something with an owner.** Documents change, policies are superseded, and a knowledge base with no maintainer converges on being a citation source for last year’s rules. The connector option exists partly to make this someone’s job by default rather than by discipline. * * * ## **Frequently asked questions** **What file types can an Insulin knowledge base hold?** PDF and DOCX, plus text formats including Markdown, plain text, CSV, TSV, logs, JSON, HTML, XML and YAML. The limit is 25 MB per file. **Can a knowledge base pull from a website or another system?** Yes. It can crawl a public website, and it connects to GitHub, ClickUp, and Google Drive with sync status visible per connector and per document. **Should I upload a file or connect the source?** Connect anything that changes. An uploaded file is a snapshot that keeps being cited long after the original was revised; a connected source stays current. **Do scanned PDFs work?** Only if they contain real text. A scan that has never been through OCR is an image to any retrieval system and will index as an empty document rather than fail visibly. **Which documents should go in first?** Something small, current, and owned — usually runbooks and approval policy. Pick material you can grade after a week, not the largest archive you have. * * * ## **Takeaways** - PDF and DOCX plus Markdown, plain text, CSV, TSV, logs, JSON, HTML, XML and YAML, at 25 MB per file. - Website crawling and GitHub, ClickUp and Google Drive connectors cover material that lives elsewhere. - Connect anything that changes; an uploaded snapshot ages silently and still gets cited. - Scanned PDFs without OCR index as empty documents rather than failing loudly. - Start with a small, current, owned source you can grade — not the biggest archive. Insulin knowledge bases index your documents and connected sources, and agents cite what they used. [Explore Insulin knowledge bases](/knowledge-bases/) or [get a demo](/schedule-demo/). ## Sources Primary sources for the platform rules cited above. Last verified August 14, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Suger Insulin docs: Knowledge Base](https://doc.suger.io/insulin/knowledge-base/) — Supported types are PDF, DOCX and text formats including .txt, .md, .csv, .tsv, .log, .json, .html, .xml and .yaml at a maximum 25 MB per file; documents come from three source types — uploaded or pasted files, crawled websites, and synced GitHub, ClickUp or Google Drive connectors. ### Stay Updated Get the latest Cloud GTM insights, product updates, and marketplace strategies delivered to your inbox.