Forum Discussion
What is the best file format for an AI agent knowledge base?
New experiment: knowledge representation vs agent instructions
After testing PDF and text-based knowledge sources, I moved to another common enterprise use case: Excel.
For a relatively large workbook with around 10 worksheets, retrieval remained inconsistent. But on a more typical business dataset — a product catalogue — Excel performed surprisingly well when rows and columns already carried clear meaning.
I then compared several representations of the same product knowledge:
| Knowledge representation | Pass rate |
|---|---|
| Structured exports | 35–45% |
| Alternative text representations | 35–45% |
| Product-centric TXT — 1 product = 1 semantic block | 67% |
Then I changed only the agent instructions:
- removed rigid response templates
- removed artificial source sections
- reduced unnecessary verbosity
- reinforced grounding and native citations
Result:
67% → 91%
Same knowledge base. Same data. Same evaluation set.
What changed in my hypothesis
I started by looking for the ideal knowledge-base format for an LLM.
I'm now seeing more of a balance.
The knowledge representation clearly matters: in this experiment, restructuring the data moved performance from roughly 45% to 67%.
But once the data was structured well enough, instruction design became an even bigger lever:
67% → 91%.
So there seems to be a sweet spot between knowledge engineering and instruction engineering.
A perfectly structured KB can still underperform if the agent is overloaded with constraints. And great instructions can only go so far when the underlying knowledge is difficult to retrieve.
My question is therefore evolving from:
“What is the best file format?” to “what is the right balance between knowledge representation, retrieval and agent instructions?”