Add pipeline for different LLMs #18
Conversation
| @@ -172,3 +172,5 @@ cython_debug/ | |||
|
|
|||
| # PyPI configuration file | |||
| .pypirc | |||
|
|
|||
| secrets/.env | |||
There was a problem hiding this comment.
Maybe we can have a catch-all for all .env files?
main.py
Outdated
| from utils.evaluation import evaluate_summary | ||
|
|
||
| def main(): | ||
| input_path = "s20e06-from-supply-chain-management-to-digital-warehousing-and-finops.md" |
There was a problem hiding this comment.
Is it possible to make this dynamic? Instead of hard coding the file name, that is.
main.py
Outdated
|
|
||
| def main(): | ||
| input_path = "s20e06-from-supply-chain-management-to-digital-warehousing-and-finops.md" | ||
| output_path = "summary_groq.md" |
There was a problem hiding this comment.
Maybe this too in case of duplicates or the user needing a different name output.
main.py
Outdated
|
|
||
| podcast_text = load_text(input_path) | ||
|
|
||
| llm = GroqLLM() # You can easily switch to OpenAI later |
There was a problem hiding this comment.
ditto. But this is me thinking out loud. Let me know your thoughts.
| 3. **Guest Introduction** | ||
| - Begin with a short introduction of the guest. | ||
| - Write one or two complete sentences. | ||
| - Each sentence should be between 25 and 40 words. |
There was a problem hiding this comment.
Isn't this requirement a bit stringent?
There was a problem hiding this comment.
You are right. But I needed something like this to achieve more consistent results that align well with Valeria’s and Ricardo’s suggestions. I have evaluated many outputs that did not match the suggested quality, and in some prompts (especially with lightweight architectures), the output was wildly changing. The final summary outputs seem much more consistent.
There was a problem hiding this comment.
By the way, I may prefer Afsin’s prompt combined with the OpenAI approach for the final case when he shares it. I’ve noticed that the output of LLMs can also vary depending on the style of the prompt we use.
| @@ -0,0 +1,10 @@ | |||
| def chunk_text(text, max_chars=3000, overlap=200): | |||
There was a problem hiding this comment.
Is chunking optional? Given most API's can handle all the context?
There was a problem hiding this comment.
This is a very good question. In my case, with two architectures I worked on (Gemma2-9B-IT and LLaMA3-70B-8192), due to token limits, I had to implement chunking. However, for Afsin’s case, it was not a problem. We may also need chunking when a longer podcast comes in for Afsin’s GPT-4o-mini model.
Description
Related Issues
Fixes #20
Changes Made
How to Test
Screenshots (if applicable)
Checklist