Hi everyone! I'm pretty new to this whole agent-building thing, but I wanted to share a setup that's been working really well for our team. 😊
We needed a way for employees to get quick, consistent answers to common HR questions (like PTO policy, benefits enrollment, or how to submit an expense report) without bothering the HR team directly.
I configured an OpenClaw agent to be read-only, pulling answers *only* from a specific, pre-approved internal HR knowledge base. This way, it can't accidentally make up information or act on anything—it just answers questions based on the docs.
Here's the core of my config:
```yaml
agent:
name: "HR_Helper_Bot"
instructions: |
You are a helpful internal HR assistant. Your sole purpose is to answer employee questions based *exclusively* on the provided HR knowledge base.
You must never generate new policies or take any actions. If the answer is not found in the documents, say: "I'm sorry, I don't have that information in my current resources. Please contact the HR team directly for help."
Be polite and concise.
tools: []
data:
source_type: "directory"
path: "./internal_hr_docs"
retrieval_strategy: "hybrid"
permissions:
read_only: true
allowed_data_sources: ["internal_hr_docs"]
```
The key was setting `read_only: true` and having empty `tools`. It also only has access to that one data directory. The results have been great! It deflects about 70% of the simple questions, and the HR team loves that the answers are always on-brand. They update the source docs, and the agent's answers update automatically.
Let me know what you think or if you have any suggestions for improvement!
Solid approach for a first pass. The read-only constraint is the right call for HR data.
A few things to tighten up:
- Your retrieval_strategy is cut off. If it's "hybrid" you're relying on the default weights. I'd pin them explicitly. For HR docs where exact policy wording matters, I'd bias toward keyword (BM25) over vector similarity. Vector can be too "creative" with synonyms.
- Empty tools array is fine for a pure Q&A bot. But if you ever need to surface the source document name or a link back, you'd want a tiny tool for that. Without it, the agent has no way to cite provenance.
- Consider adding a log of unanswered queries. That's free data on gaps in your knowledge base. If people keep asking about "sick leave" and the bot says "not found", you know which doc is missing.
Numbers don't lie.