patchfile.diff

*** a/backend/app/routers/chat.py
--- b/backend/app/routers/chat.py
***************
*** 1,20 ****
  # app/routers/chat.py
  from __future__ import annotations
  import asyncio
  from typing import List
  import re
  from fastapi import APIRouter, Depends, HTTPException
  from pydantic import BaseModel, Field
  from sqlalchemy.orm import Session
  import httpx

  from app.auth import Principal
  from app.config import settings
  from app.db import get_db
  from app.rag import search_chunks
- from app.agents import route_question, build_system_prompt
+ from app.agents import route_question, build_system_prompt
+ from app.llm_provider import llm_provider
  from app.models import Chunk, Document
***************
*** 85,102 ****
  class AskPayload(BaseModel):
      question: str
      agent: str | None = None
      top_k: int = Field(default=settings.TOP_K, ge=1, le=settings.MAX_CANDIDATE_CHUNKS)
      extended: bool = Field(default=False, description="Enable extended context by including adjacent pages")
      page_window: int = Field(default=1, ge=0, le=5)
+     # NEW: allow caller to select the LLM; accepted aliases: "gpt"|"openai" or "deepseek"
+     provider: str | None = Field(default=None, description="'gpt'|'openai' or 'deepseek'")

***************
*** 370,432 ****
      # ... previous code that prepares `messages`, `retries`, etc.

-     answer = None
-     async with httpx.AsyncClient(timeout=settings.HTTP_TIMEOUT_SECONDS) as client:
-         for attempt in range(max(1, retries)):
-             try:
-                 response = await client.post(
-                     f"{settings.CHAT_PROVIDER_BASE_URL}/chat/completions",
-                     headers=headers,
-                     json=completion_payload,
-                 )
-                 response.raise_for_status()
-                 answer = response.json()["choices"][0]["message"]["content"]
-                 break
-             except httpx.HTTPStatusError as exc:
-                 # ... existing error handling ...
-                 pass
-             except httpx.RequestError as exc:
-                 # ... existing error handling ...
-                 pass
-     if answer is None:
-         raise HTTPException(status_code=502, detail="LLM request failed after retries")
+     # --- FIX START ---
+     # Always define used_provider before any return paths, so it never triggers NameError.
+     used_provider = (payload.provider or "openai")
+     # Call provider-agnostic client; it returns {'content': ..., 'provider': ...}
+     result = await llm_provider.chat_completion(
+         messages=messages,
+         provider=payload.provider,
+         temperature=settings.CHAT_TEMPERATURE,
+         max_tokens=settings.CHAT_MAX_TOKENS,
+         retries=retries,
+     )
+     answer = result["content"]
+     # Update with the actual provider that answered (after fallback logic if any)
+     used_provider = result.get("provider", used_provider)
+     # --- FIX END ---

      answer_text = answer.strip()

      # ... code building sources/citations/agent, etc.

-     return {
+     return {
          "ok": True,
+         "provider": used_provider,
          "answer": answer_text,
          "sources": source_labels,
          "agent": resolved_agent,
      }