This week’s Sunday Coffee & Coding was less about shiny new features and more about approach. The RFP Response Agent that uses RAG to ground its answers in company material was performing ingestion inconsistently and with limited time this week, this was the only battle I could fight - on paper, the flow was simple. In practice, it was just inconsistent.
The core issue didn’t seem to be the tooling, it seemed to be responsibility overload.
I originally asked a single agent to:
- Ingest content from a single PDF file into the RAG database
- Extract requirements from another file
- Output those requirements as structured JSON
Sometimes it worked beautifully and other times the ingestion didn’t happen at all, with no obvious error.
What finally stabilized things was stepping back and doing something very un-AI-fancy - break the problem into discrete components.
I split the workflow into two agents:
- Agent 1: Ingestion only. One job. Read source material and reliably populate the RAG store.
- Agent 2: Extraction only. Consume the RFP and produce clean, hierarchical JSON.
Once each agent had a single, well-bounded responsibility, consistency improved immediately.
It was a good reminder that:
- Agents are not magic.
- Clear boundaries matter more than clever prompts.
- Decomposition still wins, even in agentic systems.
Next is fighting with the RAG content and whether I have a chunking strategy issue or not.
---------
EDIT: Just after wrapping up it occurred to me that I may have a context window size problem, not a RAG problem. Quickly increased the context window size to 128,000 tokens and I saw:
- More chunks being ingested
- More responses from the RAG query (no surprise)
- More accurate responses to the RFP question (again, no surprise)
It looks like my RAG problem was actually a context window size issue - likely only seeing a small portion of the company info document after Docling conversion.
---------
Next Step: Now the RAG portion is good enough it’s time to split the agents up and implement a multi-agent solution with A2A - fun!
GitHub Repo: https://github.com/steveh250/MAF-RFP-Factory