Downgrade conv_action_items to gpt-4.1-mini

[beastoin]

Parent: #4635

Problem

extract_action_items() uses llm_medium_experiment (gpt-5.1) but action item extraction is a well-constrained task with clear rules. gpt-4.1-mini may produce comparable quality at a fraction of the cost.

File: backend/utils/llm/conversation_processing.py:531

chain = prompt | llm_medium_experiment | action_items_parser  # gpt-5.1

Proposed Change

1. Switch extract_action_items() to use llm_mini (gpt-4.1-mini)
2. Run quality comparison on ~100 recent conversations (compare action items from both models)
3. If quality delta is acceptable (<5% difference), ship the change

Impact

• conv_action_items is 19.8% of total spend — model swap could cut this by ~80%
• Estimated 10-15% reduction in total LLM costs

Risk

Medium — action item quality may degrade for complex conversations. Needs A/B quality comparison before shipping.

Validation

Query based-hardware:llm_usage.raw for conv_action_items to get baseline token counts, then compare costs after swap.

[beastoin]

📝 The proposed solutions here are based on code analysis and have not been verified by the CTO. If you have a better approach to address this, feel free to propose it.

[aaravgarg]

gpt-4.1-mini should not be used anywhere in the app, it is almost a 1-year old model @beastoin

[beastoin]

Updated: Feb 5 BQ Data

conv_action_items confirmed as the #1 cost offender:

• Only 4.9% of total calls
• But 24.1% of total LLM cost

This is the highest cost-per-call feature in the entire system. Model downgrade from gpt-5.1 to gpt-4.1-mini would have the single biggest impact — gpt-5.1 drives 86.4% of total cost while gpt-4.1-mini handles 76.4% of calls at just 13.6% of cost.

[beastoin]

Status: Discarded per manager directive

No model downgrade — keeping gpt-5.1 for action item quality. Cost reduction pursued through flow optimization instead (see #4635 analysis).

---

by AI for @beastoin

[beastoin]

Closing — we're not pursuing model downgrades to avoid reducing output quality.