AI knowledge base from your PDFs: a guide

At three on a Friday afternoon someone in sales needs the warranty terms for one customer and opens the shared folder. Four files wait there, called contract, contract final, contract final v3 and contract final v3 new. An AI knowledge base exists for that moment, because it turns the PDFs, contracts and manuals you already have into a searchable source that Claude or ChatGPT quotes instead of guessing from.

The technique behind it is called RAG, meaning a method that looks up the relevant passages in your documents and adds them to the question before the AI answers.¹ This guide walks through the three steps from folder chaos to a working knowledge base, the documents to start with and the data questions a European company has to settle first.

Why does AI give generic answers without an AI knowledge base?

Claude and ChatGPT know a great deal about the world and nothing about your company, so a question about your returns policy gets a plausible average of everyone else's. Uploading one PDF helps for one question, while uploading fifty and asking something specific often gives an answer that sounds right and mixes up two contracts.

PDFs were designed for printing and not for machines, which is the deeper problem underneath. Tables turn into strings of numbers, columns run into each other and page headers appear in the middle of sentences, so the AI reads a scrambled version of what you see on screen.

Claude Projects soften the problem for small collections, since they switch on retrieval automatically when the files approach the context limit and can then hold up to ten times more.² That covers a team folder of a few dozen documents well, although a company-wide collection with access rights per department still needs its own setup.

Three steps from PDF to AI knowledge base

The process has three steps that sound simple, and each one has a place where it usually goes wrong.

Step 1: Extract

Your AI first has to read the documents properly, which means pulling out the text, the tables, the headings and the structure of every PDF. Extraction has to know what is a heading, what is a table and what is running text, since a clause lifted out of its section loses the meaning the section gave it.

A poor extraction makes everything after it unreliable, which is why this step takes the longest on scanned documents. Even a contract scanned at an angle in 2016 can need a person to type in the one table the scanner mangled.

Step 2: Structure

Raw text is not enough, so each document gets two kinds of output. Running text such as policies and contract clauses becomes Markdown, meaning plain text with simple marks for headings and lists. Tables, price lists and specifications become structured data that the AI can look single values up in.

The split matters because a policy on parental leave reads best as connected text, while a price list with 200 products needs rows and columns. Searching a novel and searching a reference table are different jobs, and the knowledge base does each one its own way.

Step 3: Index and verify

The structured content goes into a searchable index, and when someone asks a question the system finds the relevant passages and answers from them. Anthropic measured that adding context to each passage before indexing cut failed searches by 49 per cent, from 5.7 to 2.9 per cent, and by 67 per cent when a reranking step was added.¹

Verification is the dull part that decides whether people trust the system. Check a sample of twenty questions against the source documents before launch. The gap between "I think the return period is 14 days" and "section 4.2 of the customer contract dated 15 March 2025 says 30 days" is the whole point.

What changes when AI stops guessing?

Three situations show the difference, and each one still keeps a person in the loop.

HR: An employee asks what the policy says about parental leave for part-time staff, and instead of reading a fourteen-page PDF from 2019 the HR lead gets the answer with the paragraph number in seconds. HR still owns the policy, and an outdated PDF in the index gives an outdated answer.

Sales: A customer on the phone asks for a product specification, and the salesperson has the answer with a reference to the product sheet before the call ends. The product team still updates the sheets, since the knowledge base only knows what was indexed last.

Finance: Before renewing a supplier contract, the controller asks the knowledge base to compare payment terms and price adjustment clauses across twelve old agreements. The comparison arrives as a table in a minute instead of half a day with twelve PDFs open side by side, and the controller still reads the two clauses that decide the negotiation.

Which documents should you start with?

Start with documents that meet three conditions, because thirty well-structured files beat three hundred messy ones.

1. People ask about them often: HR policies, customer contracts and product specifications usually top the list, and a question that comes up every week marks your starting point.
2. They rarely change: A manual that holds for years suits a knowledge base far better than a policy updated every month.
3. Looking them up costs time: A document that takes twenty minutes to find five times a week costs more than eight hours a month on its own.

Leave meeting notes, personal drafts and documents under revision out of the first version, since they change too fast and contradict each other.

Your data, your rules

A knowledge base is not the same thing as staff pasting contracts into a public chat tool, which is often what happens in companies without one. Access can be set per document and per team, so the finance folder stays with finance even when everyone uses the same search.

Location is the second question, and a European company has a clear option there. Amazon Bedrock runs Claude through an EU cross-region profile that keeps requests inside the EU geography, covering regions such as Frankfurt, Ireland, Paris and Stockholm.³ GDPR still applies to whatever personal data sits in the documents, so leave personal data out of the first index unless your data protection assessment covers it.⁴

We build knowledge bases through satori-rag at a fixed price agreed after a scoping call. Teams that want to practise on their own documents first can start with satori-sprint, five one-hour sessions for €400 per user in groups of three to eight. All satori. prices are fixed and exclude VAT.

The short version

An AI knowledge base lets Claude answer from your own documents with the paragraph it used, built in three steps from extraction to verification. Start with thirty documents people ask about every week and check twenty answers before anyone relies on it.

List the five questions your team asked each other most often last month, and next to each one write the name of the document that holds the answer. In fifteen minutes you have the first five files of your knowledge base and a test set to check it against.

---

Sources

Footnotes

1. Anthropic (2024). Introducing Contextual Retrieval. https://www.anthropic.com/news/contextual-retrieval ↩ ↩²

2. Anthropic (2026). Retrieval augmented generation (RAG) for projects. Claude Help Center. https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects ↩

3. Amazon Web Services (2026). Regional availability of models in Amazon Bedrock. https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html ↩

4. European Union (2016). Regulation (EU) 2016/679, General Data Protection Regulation. https://eur-lex.europa.eu/eli/reg/2016/679/oj ↩