ai-cfia / ai-cfia/fertiscan-pipeline

As a dev, I want to explore breaking the FertilizerInspection model into smaller units for parallel LLM queries

Ouverte
#55 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub
Langage dominant
Python
Étoiles
1
Forks
1
Métriques de merge des PR
Aucune PR mergée en 30 j

Description

## Description

**Context**
In the Fertiscan pipeline, text extracted from fertilizer labels, which can vary in size, is passed to an LLM along with a full Pydantic model schema. The LLM uses these inputs to generate structured JSON from the text. Since the combined input can sometimes be large, there is a hypothesis that breaking the model into smaller parts could improve LLM performance, as it is often more efficient with smaller, more focused tasks.

**Exploration Objective**
We want to explore the potential benefits of breaking the schema into smaller, independent units and querying the LLM in parallel. This exploration will focus on whether this process improves performance, such as response time and accuracy.

## Acceptance Criteria
- Explore the feasibility and impact of breaking the model schema into smaller independent units.
- Investigate LLM performance when making parallel queries with smaller inputs versus a single large request.
- Analyze the overall response quality and ease of recombining the results into a final JSON structure.

## Additional Information
- While costs should not be a primary factor in this exploration, the potential impact on cost should still be mentioned in the analysis.

Guide de contribution

Aucun guide de contribution indexé pour ce dépôt

Évaluation

Cette issue n'a pas encore été évaluée.

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.