[EXHAUSTED_RETRIES] attempts=4 for tailored resume

[Mr-Nobody1]

`======================================================================
STAGE: tailor — Resume tailoring (LLM + validation)
Started: 00:59:29

00:59:29 - INFO - Tailoring resumes for 1 jobs (score >= 7)...
00:59:47 - INFO - HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/openai/chat/completions "HTTP/1.1 200 OK"
01:00:01 - INFO - HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/openai/chat/completions "HTTP/1.1 200 OK"
01:00:18 - INFO - HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/openai/chat/completions "HTTP/1.1 200 OK"
01:00:33 - INFO - HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/openai/chat/completions "HTTP/1.1 200 OK"
01:00:33 - INFO - 1/1 [EXHAUSTED_RETRIES] attempts=4 | 0.9 jobs/min | Intern - Generative AI Research Engineer
01:00:33 - INFO - Tailoring done in 63.6s: 0 approved, 0 failed_validation, 0 failed_judge, 0 errors

Stage 'tailor' completed in 63.6s — ok`

[KDoyle001]

im getting this error also

[Thunderhurst]

I am also seeing this error on a fresh install, both with paid and free Gemini Flash 2.5 calls and lenient validation specified. Happy to provide additional information for anyone looking to solve this.
@ibarrajo , for us newbies, could you share what you were able to do to workaround this issue?

[ibarrajo]

Hey @Thunderhurst — happy to share what worked

Root cause: The LLM calls succeed (you can see the 200 OKs) but the tailored resume fails validation every attempt. The most common culprits:

1. Gemini 2.5 Flash / thinking token budget exhaustion — Gemini 2.5 models use internal "thinking tokens" that silently consume your max_tokens
   budget, causing the actual JSON output to be truncated. If the JSON is cut off mid-field, it fails parsing and validation. Fix: increase
   max_tokens for the generation call in tailor.py to something like 16384. It is likely too low.
2. Overly strict fabrication validation — The validator sometimes flags skills that actually appear in your resume as "fabricated."
3. LLM judge blocking on the final attempt — the LLM judge (a second pass that checks for fabrication) can reject the
   resume even after validation passes.

The quickest fix: bump max_tokens to 16384 in tailor.py in the _tailor_one() function where it calls
client.chat(messages, ...).

[Thunderhurst]

@ibarrajo, thank you so much, this was incredibly helpful! I have now reached the stage where the errors have changed from the above to failing validation due to truncated tailored resumes. But this is a big step forward!
For others looking to make a similar fix, the exact lines changed were from:
response = client.chat(messages, max_tokens=512, temperature=0.1)
to 4096, and
raw = client.chat(messages, max_tokens=2048, temperature=0.4)
to 16384.
Cheers!