Preparing the knowledge base for an AI assistant
Short Answer
An AI assistant answers only as well as the sources you prepare for it. Preparing means sifting and discarding, deciding one valid version per topic, cutting documents into named sections, and adding metadata like date and origin. This work happens before the technology and shapes answer quality more than the model does.
Most knowledge bases for AI assistants fail not on the build but on the preparation that gets skipped before it. An assistant fed raw folders finds three contradicting versions, a draft from two years ago and a PDF with no recognisable structure, and forms a convincing, often wrong answer from them. Preparation is the editorial work on the sources themselves: sift what carries knowledge, decide the authoritative version per topic, remove contradictions and duplicates, cut documents sensibly and stamp them with date, validity and origin. The single strongest lever is one valid version per topic, because without it the assistant chooses between truths without seeing the conflict. This work needs no AI expertise, only someone who knows the topic.
Preparing Means Discarding
People preparing a knowledge base for an AI assistant often picture the wrong thing: gather every folder and pour it into a store. That is the opposite of preparation. It moves the chaos to a new place where it is only harder to see.
Preparation is editorial work. You go through the existing documents and decide for each one: does it carry knowledge, or does it just sit around? An approved handbook carries. A draft beside it, three mail attachments with half-finished states and the private note from someone long gone do not. To a search system, though, they all look the same, and that is exactly the problem.
A knowledge base does not get better when you put more into it. It gets better when you decide what counts and leave the rest out.
The reason is simple. An AI assistant cannot know which of five files is the right one if nobody told it. It searches for the passage closest in substance and answers from that. If three versions sit there as equals, it catches one here and another there. Discarding before the import is therefore not tidying for its own sake but the condition for the answers to be correct at all.
One Valid Version Per Topic
The single most important decision in preparation is the authoritative version. For each topic there is exactly one version that counts as true: the approved travel policy, the current approval process, the valid price list. A human who owns the topic makes this call, not a model and not the file date.
Where this decision is missing, the most expensive errors arise, because they are quiet. The assistant does not contradict itself visibly. It gives one answer today and another tomorrow, both sound clean, and nobody notices the conflict until a decision has been built on it. That is why preparation involves not only deciding the valid version but also removing or clearly marking the old one.
| Document type | Handling in preparation | Why |
|---|---|---|
| Approved policy, valid handbook | include, mark as binding | carries the authoritative answer |
| Three versions of the same process | reduce to the valid one, archive the rest | prevents contradicting answers |
| Draft, old state, private note | leave out | otherwise creates plausible noise |
| Scan, photo, table without text | extract first, then structure | otherwise invisible to retrieval |
| Personal or regulated content | only with legal review and rights | legally sensitive |
The second row is the real core of preparation. Turning five versions into one is not technical work but domain work, and it cannot be delegated to the AI. Whoever skips it builds the contradictions in for good and wonders later about shifting answers.
Structure Decides the Hit Rate
Once you have sifted and decided the valid version, comes the part that decides the hit rate: structure. A search system does not find a whole document but a passage inside it. If the valid rule sits somewhere on page 40 of a 60-page PDF, without a heading, without a clear cut, retrieval finds it poorly, even though it is there.
Preparing here means bringing documents into named, topically cut sections rather than storing them as one block. Each section ideally answers one question. Add metadata: a date, a note on validity, the origin. These stamps are what separate an answer with a citation from a guess, and they are created during preparation, not in operation.
| Question to a document | Poorly prepared | Well prepared |
|---|---|---|
| Where is the valid rule? | scattered in the long PDF | its own named section |
| Is this still current? | no date, no approval | date and validity as metadata |
| Which version counts? | three sit there as equals | one marked, the rest archived |
The effort is real, but it falls once and pays back on every answer. A cleanly cut, dated document stays findable and checkable, while a raw PDF pile grows less reliable with every new question.
How to Prepare the Sources
Before the first file moves into the knowledge base, work through these five steps. They are preparation as action, and they keep the base from becoming a mirror of the old chaos.
- Sift the sources and separate what carries knowledge from what just sits around, before you think about import.
- Decide the valid version per topic and let the responsible person name it, not the file date.
- Remove or mark old versions, so no two truths sit side by side.
- Cut documents into named sections and add date, validity and origin as metadata.
- Extract image-based sources first, because a scan or a table without text is invisible to search.
Core sentence: preparing the knowledge base is in small part technology and in large part a domain decision about what counts. Whoever skips that decision builds the errors in for good, before the assistant sees the first question.
Common Misconceptions
- More documents mean better answers.
- More documents without preparation mean more contradictions and more noise. The hit rate sinks because the search system can catch the wrong one among five similar passages. Fewer but valid sources beat many unchecked ones.
- The AI sorts the chaos out on its own.
- A model can summarise text, but it cannot know which of three versions holds true in your house. That decision is a domain call and must be made before the import, otherwise the assistant guesses convincingly.
Questions and Responses
- Must the knowledge base be complete before the assistant starts?
- No. More important than completeness is that what is in there is valid, cut and dated. A narrow, cleanly prepared area delivers better answers than a broad raw store, and you can expand after the first reliable area runs.
- How do I recognise the valid version when several circulate?
- Not by the most recent file date but by the approval. Ask who is responsible for the topic and let that person name the binding version. Where nobody decides, there is no valid version, and that is a finding you must clear before the build.
- What takes the most time in preparation?
- Usually not the cutting or dating but the deciding of what counts and what must go. This domain work cannot be automated, but it falls once and makes every later answer more reliable.
Sources and Evidence
No external statistics or client results are asserted as verified. Legal and access questions are checked against the deployed stack before connection.
