woody
/
legco_ai_assistant
1
1
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: add full input/output logging to vLLM structured output path 

Log the complete prompt, schema, extra_body content, full API response,
token counts, and full parsed JSON output. Add exc_info=True tracebacks
on all failure paths.
This commit is contained in:
Woody 2026-04-29 16:52:26 +08:00
parent 3ab6fd102a
commit 16de8394aa
1 changed files with 11 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
backend/app/services/llm_client.py
View File

@ -140,10 +140,10 @@ class LLMClient:
async def _complete_structured_vllm(self, prompt: str, pydantic_model, step_name: str, start_time: float): async def _complete_structured_vllm(self, prompt: str, pydantic_model, step_name: str, start_time: float):
"""Use vLLM's native guided_json via extra_body for structured output.""" """Use vLLM's native guided_json via extra_body for structured output."""
schema = pydantic_model.model_json_schema() schema = pydantic_model.model_json_schema()
prompt_preview = self._truncate_prompt_for_log(prompt, first_chars=300, last_chars=100)
self.logger.info( self.logger.info(
"[%s] vLLM structured: sending guided_json schema=%s", "[%s] vLLM structured: prompt=%s schema=%s",
step_name, step_name, prompt_preview, json.dumps(schema)[:300],
json.dumps(schema)[:300],
) )
# Try the new unified format first, then legacy guided_json # Try the new unified format first, then legacy guided_json
@ -152,18 +152,22 @@ class LLMClient:
("guided_json", {"guided_json": schema}), ("guided_json", {"guided_json": schema}),
]: ]:
try: try:
self.logger.info("[%s] vLLM structured: trying format=%s", step_name, fmt_name) self.logger.info("[%s] vLLM structured: trying format=%s extra=%s", step_name, fmt_name, extra)
response = await self._client.chat.completions.create( response = await self._client.chat.completions.create(
model=self.model, model=self.model,
messages=[{"role": "user", "content": prompt}], messages=[{"role": "user", "content": prompt}],
temperature=0.0, temperature=0.0,
extra_body=extra, extra_body=extra,
) )
self.logger.info("[%s] vLLM structured full response: %s", step_name, response)
content = response.choices[0].message.content or "" content = response.choices[0].message.content or ""
elapsed_ms = (time.perf_counter() - start_time) * 1000 elapsed_ms = (time.perf_counter() - start_time) * 1000
self.logger.info( self.logger.info(
"[%s] vLLM structured succeeded with format=%s in %.2fms. Response: %s", "[%s] vLLM structured succeeded format=%s in %.2fms tokens=(%s/%s). Parsed=%s",
step_name, fmt_name, elapsed_ms, content[:200], step_name, fmt_name, elapsed_ms,
response.usage.prompt_tokens if response.usage else "?",
response.usage.completion_tokens if response.usage else "?",
content,
) )
return pydantic_model.model_validate_json(content) return pydantic_model.model_validate_json(content)
except Exception as exc: except Exception as exc:
@ -171,6 +175,7 @@ class LLMClient:
self.logger.warning( self.logger.warning(
"[%s] vLLM structured format=%s failed after %.2fms: %s", "[%s] vLLM structured format=%s failed after %.2fms: %s",
step_name, fmt_name, elapsed_ms, exc, step_name, fmt_name, elapsed_ms, exc,
exc_info=True,
) )
elapsed_ms = (time.perf_counter() - start_time) * 1000 elapsed_ms = (time.perf_counter() - start_time) * 1000