My Company Knowledge Base Finally Became Actually Searchable#

"Where's that doc about the pricing change?"

Someone asks this in Slack every week. Someone else pastes a Notion link. The link is three months stale and half the page was rewritten since.

I got tired of it. So I built a different setup: a Telegram topic, an OpenClaw agent, and one skill.

Now I ask "what do we have about onboarding flows?" and get ranked answers with sources in seconds.

Here's the whole thing.

The problem was never storage. It's retrieval.#

Every team I know has the same stack: a Notion workspace, a Drive folder, and a few hundred Slack threads where decisions go to die.

And every search experience in that stack is bad in the same way.

Notion search is keyword matching. Search "why are users leaving" and it won't find the doc titled "Q2 churn analysis." Slack search is worse — you need to remember the exact channel, the exact week, and the exact phrasing someone used at 4pm on a Tuesday.

So we compensate by organizing. Folders, tags, databases, templates. It feels productive.

It isn't. The bottleneck is retrieval, not filing. You don't have a storage problem — you have a "find the thing" problem, and semantic search fixes that in a way no folder structure ever will.

That's my actual opinion after running this for a while: stop organizing, start ingesting. One pipe, everything goes in, search later. It beats any meticulously maintained wiki, mostly because wikis rot and this doesn't.

The setup: one channel, one agent, one skill#

The architecture is almost embarrassingly simple.

You create a Telegram topic called knowledge-base — or a Slack channel, if that's where your team already lives. When anyone drops a URL into it, OpenClaw fetches the content and ingests it into a knowledge base with metadata: title, URL, date, type.

It handles articles, tweets, YouTube transcripts, and PDFs. The PDF support matters more than it sounds — a huge chunk of internal knowledge lives in PDFs that nobody has opened since the quarter they were made.

Then the same topic doubles as your search interface. Ask a question, get semantic results with sources and excerpts.

Semantic, not keyword. Ask "what did we save about agent memory?" and it finds the stuff titled "long-term context for LLM apps." That's the entire value proposition.

One detail I insist on: if there are no good matches, the agent says so instead of confidently making something up. That's a prompt choice, not magic. Steal it.

Building it, step by step#

1. Install the knowledge-base skill from ClawdHub. If you'd rather roll your own, custom RAG with embeddings works too — but the skill gets you running in minutes.

2. Create the ingestion channel. A Telegram topic named knowledge-base, or a Slack channel. Pick whatever your team actually uses daily. A research repo nobody visits is just a landfill with extra steps.

3. Prompt OpenClaw. This is the entire config:

When I drop a URL in the "knowledge-base" topic:
1. Fetch the content (article, tweet, YouTube transcript, PDF)
2. Ingest it into the knowledge base with metadata (title, URL, date, type)
3. Reply with confirmation: what was ingested and chunk count

When I ask a question in this topic:
1. Search the knowledge base semantically
2. Return top results with sources and relevant excerpts
3. If no good matches, tell me

Also: when other workflows need research (e.g., video ideas, meeting prep),
automatically query the knowledge base for relevant saved content.

That's it. No pipeline code, no vector database admin, no cron jobs. The fetching runs on web_fetch, which is built in.

4. Test it. Drop in a few URLs and ask real questions.

A drop gives you a confirmation like this:

Ingested: "Why RAG systems fail in production" (article)
Source: https://example.com/rag-failures
Chunks: 24

And a question gets you ranked results with excerpts and the original URL attached:

Q: What do I have about LLM memory?

1. "Memory systems for long-running agents" — saved 2025-04-12
   "...hierarchical memory with periodic summarization holds up
   better than raw transcript storage..."
   source: https://example.com/agent-memory

Don't skip the chunk count in that confirmation prompt. It looks like a nice-to-have. It's actually your smoke test. If a 40-page PDF comes back as 3 chunks, ingestion glitched and your search results will be garbage. If it comes back as 60, you know the thing is genuinely indexed.

The part that made it stick#

A searchable archive is nice. It's not why this survived on my team past week two.

The last line of that prompt is the real feature: other workflows can query the knowledge base automatically. My video idea pipeline pulls relevant saved content when it builds research cards. Meeting prep pulls context the same way.

Once everything you save lives behind a semantic search that any workflow can call, the knowledge base stops being a graveyard and becomes shared infrastructure. Every new automation you build inherits your entire research layer for free.

That's the compounding part. Most bookmarking tools are a dead end. This is a dependency.

Honest notes before you build it#

Garbage in, garbage out. It indexes exactly what gets dropped into the channel. If your team never posts, you have an empty brain. The fix is cultural, not technical: make it the default "dump it here" spot for anything worth keeping.

One channel means discipline. The system works because there's exactly one ingestion point. The moment people start saving to three different places, you're back to Notion.

Sources are non-negotiable. Every answer comes back with the URL and an excerpt, so anyone can verify before repeating it in a meeting. This is what makes it trustworthy enough for team use — you're not asking people to trust the model, you're asking them to trust a link they can click.

Try it this week#

Total setup time is under an hour, and most of that is picking which Slack channel to sacrifice.

Install the knowledge-base skill from ClawdHub, make the channel, paste the prompt, drop in five URLs your team references constantly, then ask a question you'd normally ask in #general. The first time it answers correctly in four seconds, you'll get it.

If you want more workflows like this — practical agent setups you can ship in an afternoon, written by people who run them daily — head over to papayaclaw.com. That's where this whole playbook lives.