scriptoriumadmin/Roo-Code
1
0
Fork 0
mirror of https://github.com/RooVetGit/Roo-Code.git synced 2026-09-19 00:01:19 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
Roo-Code/IMPLEMENTATION-REPORT.md
Roo Code c5eb84b766 fix(embeddings): add timeouts, 5xx retry, and proper error messages for OpenAI Compatible embedder 
Fix critical issues where the OpenAI Compatible Embedder would hang indefinitely
on unresponsive servers and not retry 5xx server errors.

Changes:
- Add 60s timeout to OpenAI SDK constructor (timeout: 60000, maxRetries: 0)
- Add AbortController with 60s timeout to makeDirectEmbeddingRequest()
- Convert AbortError to HTTP 504 (Gateway Timeout)
- Extend retry logic to handle 5xx errors (500-599) with exponential backoff
- Update validation error messages:
  - 429 -> rateLimitExceeded
  - 502 -> badGateway (new)
  - 503 -> serviceUnavailable
  - 504 -> gatewayTimeout (new)
  - Other 5xx -> serverError (was configurationError)
- Add i18n translations for new error messages in 17 languages
- Add unit tests for timeout handling and 5xx retry (5 new tests)
- Add unit tests for getErrorMessageForStatus (10 new tests)

Impact:
- All OpenAI-compatible embedders benefit (Gemini, Mistral, VercelAiGateway, OpenRouter)
- No breaking changes - existing functionality preserved
- Prevents infinite waits on unresponsive embedding servers
- Clear error messages for 502/503/504 errors

Files changed: 21
- 2 source files (openai-compatible.ts, validation-helpers.ts)
- 17 i18n locale files (en, ru, de, es, fr, hi, id, it, ja, ko, nl, pl, pt-BR, tr, vi, zh-CN, zh-TW)
- 3 test files (openai-compatible.spec.ts, openai.spec.ts, validation-helpers.spec.ts)

Test results:
- code-index tests: 482 passed, 0 failed
- All project tests: 8210 total (8154 passed, 57 skipped, 0 failed)
- Test files: 582 (569 run, 13 skipped)
- Duration: 4m44s
2026-04-05 01:39:15 +03:00

5.6 KiB
Raw Blame History

Implementation Report: Embedding Indexing Fix

Summary

Fixed critical issues in the OpenAI Compatible Embedder that caused HTTP 503 errors and infinite waits during codebase indexing. The fix adds timeouts, retry logic for 5xx server errors, and proper error messages.

Problem

When indexing codebase through OpenAI-compatible API (http://0.0.0.0:11434/v1), the following error occurred:

Indexing partially failed: Only 780 of 2834 blocks were indexed.
Failed to process batch after 3 attempts:
Failed to create embeddings after 3 attempts: HTTP 503 - 503 status code (no body)

Root Cause: The OpenAI Compatible Embedder lacked timeouts and did not retry 5xx errors. The server could hang indefinitely — the program needed to handle such situations correctly.

Changes Made

1. Core Code Changes

src/services/code-index/embedders/openai-compatible.ts

  • Added timeout constants: OPENAI_COMPATIBLE_EMBEDDING_TIMEOUT_MS = 60000 (60s), OPENAI_COMPATIBLE_VALIDATION_TIMEOUT_MS = 30000 (30s)
  • OpenAI SDK constructor: Added timeout: 60000 and maxRetries: 0 (disabled built-in retry to use our own logic)
  • AbortController in fetch: Added AbortController with 60s timeout to makeDirectEmbeddingRequest(), converts AbortError to HTTP 504 (Gateway Timeout)
  • Retry for 5xx errors: Extended retry logic in _embedBatchWithRetries() to handle both 429 (rate limit) and 500-599 (server errors) with exponential backoff

src/services/code-index/shared/validation-helpers.ts

  • Updated getErrorMessageForStatus():
    • 429 → rateLimitExceeded (was serviceUnavailable)
    • 502 → badGateway (new)
    • 503 → serviceUnavailable (reused)
    • 504 → gatewayTimeout (new)
    • Other 5xx → serverError (was configurationError)

2. Localization (17 files)

Added 5 new i18n keys to validation section and 1 new key serverErrorRetry to root in all 17 locale files:

Language File
English src/i18n/locales/en/embeddings.json
Russian src/i18n/locales/ru/embeddings.json
German src/i18n/locales/de/embeddings.json
Spanish src/i18n/locales/es/embeddings.json
French src/i18n/locales/fr/embeddings.json
Hindi src/i18n/locales/hi/embeddings.json
Indonesian src/i18n/locales/id/embeddings.json
Italian src/i18n/locales/it/embeddings.json
Japanese src/i18n/locales/ja/embeddings.json
Korean src/i18n/locales/ko/embeddings.json
Dutch src/i18n/locales/nl/embeddings.json
Polish src/i18n/locales/pl/embeddings.json
Portuguese (BR) src/i18n/locales/pt-BR/embeddings.json
Turkish src/i18n/locales/tr/embeddings.json
Vietnamese src/i18n/locales/vi/embeddings.json
Chinese (Simplified) src/i18n/locales/zh-CN/embeddings.json
Chinese (Traditional) src/i18n/locales/zh-TW/embeddings.json

3. Tests

src/services/code-index/embedders/__tests__/openai-compatible.spec.ts

  • Updated existing test: 500 error now retries 3 times (was 1)
  • Added timeout handling describe block with 2 tests
  • Added 5xx retry handling describe block with 3 tests (502, 503, 504)

src/services/code-index/embedders/__tests__/openai.spec.ts

  • Fixed regression: Updated test expectations for new timeout/maxRetries parameters

src/services/code-index/shared/__tests__/validation-helpers.spec.ts

  • Added getErrorMessageForStatus describe block with 10 tests covering all HTTP status codes

Test Results

  • 21 test files — all passed
  • 482 tests — 0 failed, 0 errors, 0 warnings
  • Duration: ~9-11s

Files Changed (21 total)

File Changes
src/services/code-index/embedders/openai-compatible.ts Steps 1-4: Timeouts, AbortController, 5xx retry
src/services/code-index/shared/validation-helpers.ts Step 5: 5xx error messages
src/i18n/locales/*/embeddings.json (17 files) Step 6: i18n keys
src/services/code-index/embedders/__tests__/openai-compatible.spec.ts Steps 7-8: New tests
src/services/code-index/embedders/__tests__/openai.spec.ts Regression fix
src/services/code-index/shared/__tests__/validation-helpers.spec.ts Step 9: New tests

Architecture

Request → {Full URL?} → Yes → makeDirectEmbeddingRequest (AbortController 60s)
                    → No  → OpenAI SDK (timeout 60s, maxRetries 0)
                              ↓
                    Error? → {429 or 5xx?} → Yes → Retry with exponential backoff
                                    → No  → Throw immediately

Impact

  • All OpenAI-compatible embedders benefit: Gemini, Mistral, VercelAiGateway, OpenRouter
  • No breaking changes: Existing functionality preserved
  • Better user experience: Clear error messages for 502/503/504 errors
  • Prevents infinite waits: 60s timeout on all embedding requests