🧩 Tight coupling in LLM pipeline is causing repeated merge conflicts
Summary
While working on a few improvements around the LLM pipeline (validation layer, fallback handling, prompt consistency), I kept running into the same issue — a lot of the core logic is tightly coupled across a few files:
src/llm.pyapi/routes/forms.pysrc/main.py
Because of this, even unrelated changes end up touching the same code paths, which leads to frequent merge conflicts.
What I’m noticing
Right now, multiple responsibilities are mixed together in the same flow:
- prompt building
- calling the LLM (Ollama)
- parsing / cleaning responses
- validation
- fallback / error handling
Most of this lives inside or around LLM.main_loop().
Why this becomes a problem
- Different PRs (validation, fallback, prompt tweaks, etc.) all modify the same logic
- Conflicts keep reappearing even after resolving them once
- It’s hard to work on one concern without affecting others
- Slows down contribution and increases the chance of introducing bugs during merges
Example
While working on:
all of them end up overlapping in the same execution path, which makes rebasing messy.
Possible direction
Maybe we can move toward a slightly more modular flow like:
Input → Prompt Builder → LLM Client → Response Parser → Validator → Output
Not suggesting a big rewrite — even small separations could help, like:
- isolating prompt building
- separating the LLM call layer
- keeping validation independent (pluggable)
- reducing logic inside the main loop
Why this could help
- Fewer merge conflicts across active PRs
- Easier to work on features independently
- Cleaner structure for future additions (confidence scoring, retries, async, etc.)
- Better testability
Context
This came up while actively working on multiple PRs touching the same pipeline — I had to rebase multiple times due to overlapping changes.
Proposed approach
Rather than a large refactor, this could be tackled incrementally with small, focused PRs (e.g., extracting the LLM client, isolating prompt building, introducing validation as a separate step).
Happy to start with a small initial PR and iterate.
🧩 Tight coupling in LLM pipeline is causing repeated merge conflicts
Summary
While working on a few improvements around the LLM pipeline (validation layer, fallback handling, prompt consistency), I kept running into the same issue — a lot of the core logic is tightly coupled across a few files:
src/llm.pyapi/routes/forms.pysrc/main.pyBecause of this, even unrelated changes end up touching the same code paths, which leads to frequent merge conflicts.
What I’m noticing
Right now, multiple responsibilities are mixed together in the same flow:
Most of this lives inside or around
LLM.main_loop().Why this becomes a problem
Example
While working on:
all of them end up overlapping in the same execution path, which makes rebasing messy.
Possible direction
Maybe we can move toward a slightly more modular flow like:
Not suggesting a big rewrite — even small separations could help, like:
Why this could help
Context
This came up while actively working on multiple PRs touching the same pipeline — I had to rebase multiple times due to overlapping changes.
Proposed approach
Rather than a large refactor, this could be tackled incrementally with small, focused PRs (e.g., extracting the LLM client, isolating prompt building, introducing validation as a separate step).
Happy to start with a small initial PR and iterate.