Keep your knowledge base current

The first version is easy and it works. You extract your documents, index them, point the agent at the index, and the answers are accurate because the content is fresh.

Three months later the product docs have been rewritten, two support articles changed, a policy was updated, and a new whitepaper went out. The agent knows none of this. It answers from the version of the world it was handed at ingestion, and the gap widens every week until a customer runs into it. The failure is quiet, which is the worst part: the agent still sounds confident, it is simply wrong, and you find out from the person you least wanted to hear it from.

This page is for teams whose agent draws on a corpus that changes: product documentation, internal wikis, support content, research libraries, regulatory filings.

What you have probably already tried

The interim fixes share a failure mode, which is that they depend on somebody remembering.

Manual re-ingestion works exactly as well as your team's spare capacity. Someone exports the updated docs, runs the pipeline, replaces the old content. Then a busy month arrives, the job slips, and the knowledge base drifts without anyone noticing, because nothing breaks visibly when content goes stale.

A cron job that re-ingests everything is better and blunt. It reprocesses your entire corpus whether or not anything changed, which is wasteful at small scale and impractical at large, and it still leaves you owning the scheduling infrastructure, the failure handling, and the alerting for when it stops running.

Watching the source for changes is the right instinct and turns into a project. You need change detection per source, a queue, retry logic, and somewhere to put the re-extracted output, and now you maintain a pipeline whose only job is keeping another system honest.

Just telling users the content might be out of date is an admission rather than a fix, and it undermines the reason they were asking the agent instead of searching the docs themselves.

The underlying problem is that ingestion was built as an event when the corpus is a moving target. A knowledge base that is only as current as the last time someone ran something will always drift, and the drift is invisible until it is expensive.

How Exabase handles this

Workers are autonomous agents that run inside a Base on a schedule. You describe the task in natural language, set a frequency, and they run without being triggered.

For a knowledge base that maintains itself, the work is: re-extract documents that have changed, process and tag new material as it arrives, move resources into the right place, prune or flag content that has aged out, and update memories as the underlying material changes. That is the upkeep somebody was previously doing by hand, or not doing.

The flow underneath is straightforward. Documents live as Resources in a Base. Extract processes them into searchable content. A Worker re-extracts on a schedule so updated documents are reprocessed automatically, new content is tagged and organised, and stale material is identified. Everything stays indexed for Deep Search throughout, so the search index reflects the current state of your content rather than a snapshot from setup day.

There is no cron infrastructure to run, because Workers run inside Exabase. They also operate only within their own Base, so if you are multi-tenant, each customer's knowledge base is maintained independently without any of them touching another. That combination, per-tenant knowledge bases that each keep themselves current, is the part that is genuinely tedious to build yourself.

What changes

The knowledge base compounds instead of decaying. New documents get ingested, existing ones get refreshed, and what the agent answers from is the current state of your content rather than a historical snapshot.

The more useful change is that staying current stops depending on attention. Nobody has to remember, notice, or find time, which matters because the failure mode here was never that teams did not care about accuracy. It was that silent drift does not announce itself and manual upkeep loses to whatever is loudest that week.

If the accuracy problem is not just staleness but retrieval quality, making your RAG pipeline accurate covers that side. If the corpus needs building in the first place, extracting structured data from any source is the ingestion half.

Get started

How to build a company knowledge base for AI agents is the practical guide. The Workers page has the API detail, self-maintaining knowledge bases is the end-to-end pattern, and the topic researcher example is a working Worker-driven build to read. There is a free tier to build against.

FAQs

How does the knowledge base know a document changed?

A Worker checks the sources it is responsible for on its schedule and re-extracts the ones that have been updated, so the indexed content is replaced rather than left as the original version.

How often does it run?

On whatever schedule you set, matched to how fast your content moves. Daily suits documentation that changes often, weekly or monthly is usually enough for slower corpora.

Do I need to run cron jobs or separate infrastructure?

No. Workers run inside Exabase on the schedule you define, so there is no scheduler to operate, no retry logic to write, and no alerting to build for when your own job silently stops.

What happens to content that is no longer relevant?

Workers can identify and prune or flag outdated material, so the corpus stays clean rather than growing noisier with every addition. That matters for accuracy as much as for tidiness, since stale documents that remain indexed are exactly what produces confident wrong answers.

Can each customer have their own maintained knowledge base?

Yes. Workers operate only within their own Base, so in a multi-tenant product each customer's knowledge base is maintained independently with no crossover. Shipping multi-tenant covers the isolation model.

Does keeping content current fix inaccurate answers?

It fixes the subset caused by staleness, which is a large one. If answers are wrong because retrieval returns topically related but unhelpful passages, that is a different problem, covered in making your RAG pipeline accurate.

What formats can it re-process?

Anything Extract handles, which is PDFs, images, audio, video, Word documents, and web pages through one endpoint. Extracting structured data from any source covers the ingestion side.