Adding a dynamic RAG system to handle build errors #1216
Conversation
|
/gcbrun exp -n wentao -m vertex_ai_gemini-2-5-pro-chat -ag -b quick-test -ns 1 -- --use-error-memory |
|
Requested GKE Job: ofg-pr-1216-wentao JOB: https://console.cloud.google.com/kubernetes/job/us-central1-c/llm-experiment/default/ofg-pr-1216-wentao |
|
/gcbrun exp -n wentao -m vertex_ai_gemini-2-5-pro-chat -ag -b quick-test -rd -ns 1 -- --use-error-memory |
|
Requested GKE Job: ofg-pr-1216-wentao JOB: https://console.cloud.google.com/kubernetes/job/us-central1-c/llm-experiment/default/ofg-pr-1216-wentao |
|
/gcbrun exp -n wentao -m vertex_ai_gemini-2-5-pro-chat -ag -b minor-for-ci -rd -ns 1 -- --use-error-memory |
|
Requested GKE Job: ofg-pr-1216-wentao JOB: https://console.cloud.google.com/kubernetes/job/us-central1-c/llm-experiment/default/ofg-pr-1216-wentao |
2 participants
This PR introduces a RAG-based system to handle build errors more effectively. By retrieving previously successful fix snippets ("patch texts") from a database, the agent can reuse past solutions to avoid recurring build issues.
Implementation details:
Under
agent, we add amemory_prototyperagent that extends the general workflow of prototyper, but injects historical fix data into the context when compilation fails.Infrastructure: We add the
memory_helpermodule with utilities to access and connect to cloudSQL instance.Current Status and Purpose: This PR is a draft to verify that the memory_helper utilities can successfully authenticate and connect to the Cloud SQL instance from the CI/CD environment. It is not ready for review.