src/experience/service.py (1)

647-650: offset 기반 과다 fetch를 제한하는 방어 로직을 권장합니다.

현재 구조는 offset + limit만큼 선조회 후 Python 슬라이싱이라, 큰 offset 요청에서 메모리/응답시간이 급격히 증가할 수 있습니다. 페이지 윈도우 상한을 두는 편이 안전합니다.

제안 수정안

+        MAX_PAGE_WINDOW = 2000
+        if (limit + offset) > MAX_PAGE_WINDOW:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail=f"Pagination window too large. limit+offset must be <= {MAX_PAGE_WINDOW}.",
+            )
+
         fetch_limit = min(limit + offset, total)

Also applies to: 671-673

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/experience/service.py` around lines 647 - 650, Introduce a configurable
page-window cap and apply it to the fetch sizing logic: define a PAGE_WINDOW_MAX
(or PAGE_WINDOW_LIMIT) constant and change both usages of fetch_limit (the
current calculation using fetch_limit = min(limit + offset, total)) to include
this cap — e.g., fetch_limit = min(limit + offset, total, PAGE_WINDOW_MAX). Also
add an early guard that handles excessively large offset values (either clamp
offset = min(offset, PAGE_WINDOW_MAX - limit) or return a validation error) so
callers cannot force a huge in-memory slice; update both occurrences (the
fetch_limit assignment around fetch_limit = min(limit + offset, total) and the
similar block at lines ~671-673) accordingly.

src/chats/llm_service.py (3)

337-338: asyncio.create_task 참조를 저장하지 않으면 예외가 손실될 수 있음

태스크 참조를 저장하지 않으면 태스크 내에서 발생한 예외가 "Task exception was never retrieved" 경고만 남기고 조용히 무시될 수 있습니다. 현재 구조에서는 예외가 queue를 통해 전달되므로 큰 문제는 아니지만, 명시적으로 태스크를 추적하는 것이 더 안전합니다.

♻️ 제안: 태스크 참조 저장 및 예외 처리

- asyncio.create_task(_run())
+ task = asyncio.create_task(_run())
+ task.add_done_callback(lambda t: t.exception() if t.done() and not t.cancelled() else None)

또는 더 간단하게 태스크 이름을 지정하여 디버깅을 용이하게:

- asyncio.create_task(_run())
+ asyncio.create_task(_run(), name=f"ai_response_stream_{question_id}")

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/chats/llm_service.py` around lines 337 - 338, The background pipeline
task created with asyncio.create_task(_run()) should be stored and its
exceptions handled; change the call to save the Task (e.g., self._pipeline_task
= asyncio.create_task(self._run(), name="llm_pipeline") or append to a tasks
list) and attach a done-callback that checks task.exception() and logs/handles
errors (or await it in shutdown). Locate the _run() invocation and replace the
fire-and-forget call with storing the Task reference and adding a
task.add_done_callback(lambda t: log and re-raise or log t.exception()) so
exceptions aren't lost and debugging is easier.

---

148-154: 예외를 조용히 무시하면 디버깅이 어려워질 수 있음

get_experiences_by_ids에서 예외 발생 시 continue로 넘어가면서 로깅이 없습니다. 일부 경험을 가져오지 못해도 계속 진행하는 것은 합리적이지만, 최소한 warning 로그를 남기는 것이 좋습니다.

🔧 제안: 예외 로깅 추가

  for exp_id in experience_ids:
      try:
          exp = get_experience(qdrant_client, exp_id, user_id)
          experiences.append(exp)
-     except Exception:
+     except Exception as e:
+         logger.warning(f"Failed to fetch experience {exp_id}: {e}")
          continue

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/chats/llm_service.py` around lines 148 - 154, In get_experiences_by_ids,
don’t silently swallow exceptions in the loop over experience_ids; instead catch
the exception from get_experience(qdrant_client, exp_id, user_id) into a
variable and emit a warning log that includes the exp_id and the exception
details (and stack trace if your logger supports it) before continuing to the
next id so missing experiences are visible in logs; update the try/except block
around get_experience to call the existing logger (or create one named logger)
and include context like exp_id and user_id in the log message while still
appending successful exp results to experiences.

---

99-104: mock_scores가 하드코딩되어 있음 - 기술 부채

mock_scores = [1.0, 0.7, 0.4]는 실제 적합도 점수가 아닌 고정된 mock 값입니다. 현재는 경험의 순서에 따라 임의의 점수가 할당되어 프롬프트에 표시됩니다. 실제 적합도 점수를 계산하거나, mock 값임을 명시적으로 표시하거나, 점수 표시를 제거하는 것을 고려해 주세요.

♻️ 옵션 1: 점수 표시 제거

- mock_scores = [1.0, 0.7, 0.4]
  parts = []
  for i, exp in enumerate(experiences):
      is_primary = (i == 0)
-     score = mock_scores[min(i, len(mock_scores) - 1)]
-     parts.append(_format_experience_with_role(exp, is_primary, score))
+     parts.append(_format_experience_with_role(exp, is_primary))

그리고 _format_experience_with_role에서 score 파라미터 제거:

- def _format_experience_with_role(exp: Experience, is_primary: bool, score: float) -> str:
+ def _format_experience_with_role(exp: Experience, is_primary: bool) -> str:
      role_tag = "[주 경험 - 상세 서술]" if is_primary else "[보조 경험 - 간략 언급]"
      lines = [
-         f"### {exp.title} {role_tag} (적합도: {score:.2f})",
+         f"### {exp.title} {role_tag}",

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/chats/llm_service.py` around lines 99 - 104, The code currently assigns
hardcoded mock_scores and passes them into _format_experience_with_role, which
embeds fake suitability scores into prompts; remove this technical debt by
eliminating score handling: delete mock_scores and the score argument when
building parts in the loop (where experiences and parts are used), then update
_format_experience_with_role to remove the score parameter and any score
rendering logic, and adjust all other callers of _format_experience_with_role
accordingly; alternatively, if you prefer to surface scores, replace mock_scores
with a real scoring function (e.g., compute_experience_score(experience)) and
call that instead, ensuring the signature changes are applied consistently.

src/chats/dependencies.py (1)

54-65: OPENAI_TEMPERATURE를 Anthropic LLM에 사용하는 것은 혼란을 줄 수 있음

settings.OPENAI_TEMPERATURE를 Claude 모델에 사용하고 있습니다. 기능적으로는 동작하지만, 설정 이름이 오해를 불러일으킬 수 있습니다. 범용적인 이름(예: LLM_TEMPERATURE 또는 DEFAULT_TEMPERATURE)으로 변경하거나, Anthropic 전용 설정을 추가하는 것을 고려해 주세요.

♻️ 제안: 범용 temperature 설정 사용

src/config.py에서:

- OPENAI_TEMPERATURE: float = 0.7
+ LLM_TEMPERATURE: float = 0.7  # 또는 OPENAI_TEMPERATURE와 ANTHROPIC_TEMPERATURE 분리

그리고 이 파일에서:

  self._writing_llm = ChatAnthropic(
      model=settings.ANTHROPIC_MODEL,
      api_key=settings.ANTHROPIC_API_KEY,
-     temperature=settings.OPENAI_TEMPERATURE,
+     temperature=settings.LLM_TEMPERATURE,
  )

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/chats/dependencies.py` around lines 54 - 65, The property writing_llm
currently uses settings.OPENAI_TEMPERATURE which is misleading for an Anthropic
model; update configuration to expose a clear temperature setting (either a
generic LLM_TEMPERATURE/DEFAULT_TEMPERATURE or an Anthropic-specific
ANTHROPIC_TEMPERATURE) in src/config.py and replace settings.OPENAI_TEMPERATURE
with the new setting in the writing_llm property (and any other places that
reference OPENAI_TEMPERATURE) so the code reads ChatAnthropic(...,
temperature=settings.LLM_TEMPERATURE) or
temperature=settings.ANTHROPIC_TEMPERATURE; ensure the new config key is
documented/defaulted and update imports/typing if needed.

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro