Last weekโs dry run proved the PoC worked. A buyer could issue an RFP, a supplier could upload their company knowledge, the responder agent could generate a grounded draft response, and the assessor agent could evaluate that response against weighted criteria with rationale and evidence.
๐ฆ๐ผ ๐๐ต๐ถ๐ ๐๐ฒ๐ฒ๐ธ๐ฒ๐ป๐ฑ ๐ ๐ณ๐ผ๐ฐ๐๐๐ฒ๐ฑ ๐ผ๐ป ๐ฟ๐ฒ๐๐ฝ๐ผ๐ป๐๐ฒ ๐พ๐๐ฎ๐น๐ถ๐๐.
๐๐ฉ๐ฆ ๐ง๐ช๐ณ๐ด๐ต ๐ต๐ฉ๐ช๐ฏ๐จ ๐ ๐ค๐ฉ๐ข๐ฏ๐จ๐ฆ๐ฅ ๐ธ๐ข๐ด ๐ต๐ฉ๐ฆ ๐ฑ๐ณ๐ฐ๐ฎ๐ฑ๐ต. The original prompt was a useful starting point, but it was too broad. It produced something plausible, but not something I would want to hand to a proposal team as a first draft. I reworked it to make the responder agent pay much closer attention to the specific RFP requirement, use the supporting RFP information properly, and produce something closer to a proposal section than a general capability statement (๐ข๐ญ๐ญ ๐ฃ๐ข๐ด๐ฆ๐ฅ ๐ฐ๐ฏ ๐ฎ๐บ ๐ด๐ต๐ข๐ฏ๐ฅ๐ข๐ณ๐ฅ ๐ฑ๐ณ๐ฐ๐ฎ๐ฑ๐ต๐ช๐ฏ๐จ ๐ต๐ฆ๐ฎ๐ฑ๐ญ๐ข๐ต๐ฆ).
That alone helped. The improved prompt produced a clearer and more focused response in about 52 seconds.
๐๐ฉ๐ฆ๐ฏ ๐ ๐ธ๐ข๐ฏ๐ต๐ฆ๐ฅ ๐ต๐ฐ ๐ต๐ฆ๐ด๐ต ๐ต๐ฉ๐ฆ ๐ฆ๐ง๐ง๐ฆ๐ค๐ต ๐ฐ๐ง ๐ข ๐ฎ๐ฐ๐ฅ๐ฆ๐ญ ๐ค๐ฉ๐ข๐ฏ๐จ๐ฆ ๐ฎ๐ฐ๐ฅ๐ฆ๐ญ. I updated Ollama and moved from llama3.1:8b to Google gemma4:26b.
The full sample prompt is over 36,000 characters, not that big. The gemma4 model just spun. A simple Harry Potter prompt worked fine, so the model was not completely broken. Testing via ๐ฐ๐ญ๐ญ๐ข๐ฎ๐ข ๐ณ๐ถ๐ฏ, it just dropped back to the prompt. Tested a smaller prompt which worked and Iฬฒ ฬฒdฬฒoฬฒnฬฒโฬฒtฬฒ ฬฒkฬฒnฬฒoฬฒwฬฒ ฬฒwฬฒhฬฒyฬฒ ฬฒIฬฒ ฬฒdฬฒiฬฒdฬฒnฬฒโฬฒtฬฒ ฬฒcฬฒlฬฒuฬฒeฬฒ ฬฒiฬฒnฬฒ ฬฒoฬฒnฬฒ ฬฒtฬฒhฬฒiฬฒsฬฒ ฬฒ-ฬฒ ฬฒIฬฒ ฬฒhฬฒaฬฒvฬฒeฬฒ ฬฒcฬฒoฬฒmฬฒeฬฒ ฬฒaฬฒcฬฒrฬฒoฬฒsฬฒsฬฒ ฬฒtฬฒhฬฒeฬฒ ฬฒpฬฒrฬฒoฬฒbฬฒlฬฒeฬฒmฬฒ ฬฒbฬฒeฬฒfฬฒoฬฒrฬฒeฬฒ.ฬฒ
So I watched the Ollama logs, ๐ซ๐ฐ๐ถ๐ณ๐ฏ๐ข๐ญ๐ค๐ต๐ญ -๐ง -๐ถ ๐ฐ๐ญ๐ญ๐ข๐ฎ๐ข and there it was, ๐ต๐ณ๐ถ๐ฏ๐ค๐ข๐ต๐ช๐ฏ๐จ ๐ช๐ฏ๐ฑ๐ถ๐ต ๐ฑ๐ณ๐ฐ๐ฎ๐ฑ๐ต ๐ญ๐ช๐ฎ๐ช๐ต=4096 - the ๐ฏ๐ถ๐ฎ_๐ค๐ต๐น problem. Quick fix to the API code, a re-test and all was fine.
The response from the gemma4:26b was much better, followed the requirement more closely, used the RFP context more effectively, and read much more like something a human could refine rather than rewrite. (I also like how the Gemma4 model shows itโs reasoning - something I have captured in the past as a form of ๐ณ๐ฆ๐ข๐ด๐ฐ๐ฏ๐ช๐ฏ๐จ ๐ข๐ถ๐ฅ๐ช๐ต ๐ต๐ณ๐ข๐ช๐ญ).
Still a POC. Still rough around the edges but ๐๐๐ถ๐ป๐ด ๐ฎ ๐ฐ๐ผ๐บ๐ฝ๐น๐ฒ๐๐ฒ๐น๐ ๐ผ๐ณ๐ณ๐น๐ถ๐ป๐ฒ ๐ฏ๐๐ถ๐น๐ฑ, ๐ฎ ๐ฐ๐ผ๐ฟ๐ฟ๐ฒ๐ฐ๐๐น๐ ๐ฐ๐ผ๐ป๐ณ๐ถ๐ด๐๐ฟ๐ฒ๐ฑ ๐บ๐ผ๐ฑ๐ฒ๐น, ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ ๐ฎ๐ป ๐ฅ๐๐ฃ ๐บ๐ฎ๐ฟ๐ธ๐ฒ๐๐ฝ๐น๐ฎ๐ฐ๐ฒ ๐๐ผ ๐ฏ๐ฟ๐ถ๐ป๐ด ๐ฏ๐๐๐ฒ๐ฟ๐ ๐ฎ๐ป๐ฑ ๐๐ฒ๐น๐น๐ฒ๐ฟ๐ ๐๐ผ๐ด๐ฒ๐๐ต๐ฒ๐ฟ ๐๐ผ ๐ฟ๐ฒ๐ฑ๐๐ฐ๐ฒ ๐ณ๐ฟ๐ถ๐ฐ๐๐ถ๐ผ๐ป ๐ฎ๐ป๐ฑ ๐ผ๐ฝ๐๐ถ๐บ๐ถ๐๐ฒ ๐๐ต๐ฒ ๐ฒ๐ป๐ฑ ๐๐ผ ๐ฒ๐ป๐ฑ ๐ฝ๐ฟ๐ผ๐ฐ๐ฒ๐๐ ๐น๐ผ๐ผ๐ธ๐ ๐๐ผ ๐บ๐ฒ ๐๐ผ ๐ฏ๐ฒ ๐ฎ ๐ฟ๐ฒ๐ฎ๐น๐ถ๐๐.