Hacktoberfest 2026 : les issues que les mainteneurs ont marquées pour octobre, ouvertes et accessibles aux débutants. Parcourir les issues Hacktoberfest

Evaluation: Improve guardrails awareness

Fermée
  • Quelque chose ne va pas ?
#1,246 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

@AkhileshNegi y travaille déjà.

Depuis le 28/9/2026.

  • #1208 par @AkhileshNegi — ouverte

Évaluation

Difficulté
5/5
Temps estimé
Plus d'une semaine
Accessibilité débutants
15/100
Type d'issue
Fonctionnalité
Clarté
Clairement spécifiée
Activité
À l'abandon
Stack technique
fastapi, python
Domaine
ai, backend-api-design

Piste de recherche

The issue says the evaluation job's v2 path and GET /api/v1/evaluations/{id} results are involved; start by tracing that evaluation path and the execute_llm_call path. Check how guardrail data passes through retries, merging, scoring, and traces. Done means fast/v2 evaluations apply input and output guardrails and expose the requested per-row fields without recording eval traffic as llm_call rows; the issue reports this work as implemented in #1208.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

enhancement

Is your feature request related to a problem?
Text evaluations (fast run mode / v2) currently ignore the stored input and output guardrails. This leads to eval scores that don't accurately reflect the system's behavior in production.

Describe the solution you'd like

  • Make fast/v2 evaluations guardrails-aware: check each dataset query against input_guardrails before the LLM call and the response against output_guardrails after.
  • Use the execute_llm_call path instead of a separate eval-only path, avoiding the recording of eval traffic as llm_call rows.
  • Include per-row guardrail status, input_to_llm, and output_from_llm in GET /api/v1/evaluations/{id} results.
  • Carry guardrail data through retry, merge, scoring, and traces.
Original issue

Is your feature request related to a problem?
Text evaluations (fast run mode / v2) ignore the input and output guardrails stored on the config. The golden dataset query is sent straight to the LLM and the generated response is never checked by output guardrails, so eval scores don't reflect how the config actually behaves in production via /llm/call.

Describe the solution you'd like

  • Make the fast/v2 evaluation run guardrails-aware: run each dataset query through the config's input_guardrails before the LLM call, and the response through output_guardrails after.
  • Reuse the execute_llm_call path instead of a separate eval-only LLM path, without recording eval traffic as llm_call rows / production metrics.
  • Surface per-row guardrail status, input_to_llm and output_from_llm in GET /api/v1/evaluations/{id} results.
  • Carry guardrail data through retry, merge, scoring and traces.

Original issue

Earlier doing text evals v2 used not to take into account the stored input and output validators ids and instead sending the input / output straight to openAI. Now the evaluation job (v2) is guardrails aware; meaning the query in the golden dataset is passed through input guardrails if they are stored; similar functionality for responses generated (if o/p guardrails are configured).

Comments: Implemented in #1208.

Langage dominant
Python
Étoiles
18
Forks
11
Merge moyen
4 j 16 h
PR mergées (30 j)
11

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de ProjectTech4DevAI/kaapi-backend

Toutes les issues de ProjectTech4DevAI/kaapi-backend

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.