[Feature]: ability to fine tune the prompt context

[laurentperez]

Description

hello

context is : ollama + llama3.1 8B

the generated commit message is too rich, it acts as is the LLM did a code review of my code and gave me some suggestions.

example :

Suggested Solution

I'd like to drive the prompt as in "you are not a code reviewer. do not give me suggestions, instead tell me what I did and return a concise report of the changes, to get a compact git commit message"

how would you do this ?

Alternatives

No response

Additional Context

No response

[tanc]

Looking at the code here and here it seems like it should be providing a decent prompt but I get the same kind of result as @laurentperez does from ollama + llama3.1 or ollama + mistral-nemo.

│
◇  📝 Commit message generated
│
└  Generated commit message:
——————————————————
It seems like you're adding a new field called `memberOnly` to your GraphQL schema for the `ScientificArticle` type. Here's how the changes you've made affect the schema:

[tanc]

I just tried with fewer files staged and the result was much better. Is this a tokens issue? I've upped the number of max tokens in OCO_TOKENS_MAX_INPUT and OCO_TOKENS_MAX_INPUT but the problem remains

[tanc]

ok, this due to Ollama having a default context size to 2048. This can be increased when making API calls but this project doesn't do that. A workaround is to create a model with a larger context by doing the following (example for llama3.1):

1. Create a Modelfile text file somewhere with the following content:

FROM llama3.1

PARAMETER num_ctx 8192
PARAMETER num_predict -1

2. Run the command ollama create llama3.1-max-context -f ./Modelfile
3. You should see the new model in the list when running ollama list
4. Change the model used by oco: oco config set OCO_MODEL='llama3.1-max-context'

This should allow for much bigger diffs to be pushed into the model to get a correct answer.