Update app.py

app.py

@@ -1,6 +1,6 @@
 """
 Amazon Trailer Inspector — app.py
-HuggingFace Spaces · FastAPI · Free vision LLMs
+HuggingFace Spaces · FastAPI · Google Gemini Vision API
 
 REST API that accepts 6 labeled images and runs all 6 aspect inspections
 in parallel, returning a structured JSON inspection report.
@@ -17,31 +17,26 @@ import re
 import traceback
 from typing import Optional
 
+import requests
 import uvicorn
 from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import JSONResponse
 from PIL import Image
-from huggingface_hub import InferenceClient
 from pydantic import BaseModel, Field
 
 # ──────────────────────────────────────────────────────────────────────────────
-# MODELS  (tried in order — first success wins per image)
+# GEMINI MODELS  (tried in order — first success wins)
 # ──────────────────────────────────────────────────────────────────────────────
 MODELS = [
-    # ── Tier 1: best vision quality ──────────────────────────────────────────
-    "google/gemma-4-27b-it",             # Primary — Gemma 4 27B (stable HF serverless name)
-    "meta-llama/Llama-4-Scout-17B-16E-Instruct",  # Llama 4 Scout — excellent vision, free HF
-    # ── Tier 2: dedicated vision models ──────────────────────────────────────
-    "Qwen/Qwen2.5-VL-7B-Instruct",      # Qwen 2.5 VL — strong free-tier vision
-    "Qwen/Qwen2-VL-7B-Instruct",        # Qwen 2 VL — previous gen, very stable
-    # ── Tier 3: additional fallbacks ─────────────────────────────────────────
-    "meta-llama/Llama-3.2-11B-Vision-Instruct",   # Llama 3.2 11B Vision — reliable fallback
-    "microsoft/Phi-3.5-vision-instruct", # Phi-3.5 Vision — lightweight, good accuracy
-    "HuggingFaceM4/idefics3-8b-llama3",  # IDEFICS3 — HF native, always available
-    "mistralai/Pixtral-12B-2409",        # Pixtral 12B — Mistral's vision model, free tier
+    "gemini-2.0-flash",    # Primary — best quality, fast, free tier (1500 req/day)
+    "gemini-1.5-flash",    # Fallback — previous gen, very stable free tier
+    "gemini-1.5-flash-8b", # Fallback 2 — lightest model, always available
 ]
 
+# Gemini API base URL
+GEMINI_API_BASE = "https://generativelanguage.googleapis.com/v1beta/models"
+
 # ──────────────────────────────────────────────────────────────────────────────
 # ASPECT PROMPTS
 # ──────────────────────────────────────────────────────────────────────────────
@@ -373,16 +368,53 @@ LABEL_TO_ASPECT = {
 }
 
 # ──────────────────────────────────────────────────────────────────────────────
-# HF CLIENT
+# GEMINI API CALL
 # ──────────────────────────────────────────────────────────────────────────────
 
-_hf_client: InferenceClient | None = None
+def call_gemini(b64_image: str, prompt: str, model: str, api_key: str) -> str:
+    """
+    Call Google Gemini vision API.
+    Returns the raw text response from the model.
+    Raises requests.HTTPError on API errors.
+    """
+    url = f"{GEMINI_API_BASE}/{model}:generateContent?key={api_key}"
+
+    payload = {
+        "system_instruction": {
+            "parts": [{
+                "text": (
+                    "You are a JSON-only API for trailer inspection. "
+                    "You MUST respond with a single valid flat JSON object and absolutely "
+                    "nothing else — no explanation, no preamble, no markdown fences, "
+                    "no reasoning text, no nested objects. "
+                    "Every value must be exactly the string \"detected\" or \"missing\". "
+                    "Start your response with '{' and end with '}'."
+                )
+            }]
+        },
+        "contents": [{
+            "parts": [
+                {
+                    "inline_data": {
+                        "mime_type": "image/jpeg",
+                        "data": b64_image,
+                    }
+                },
+                {
+                    "text": prompt,
+                }
+            ]
+        }],
+        "generationConfig": {
+            "temperature":     0.05,
+            "maxOutputTokens": 120,
+        },
+    }
 
-def _get_client(token: str) -> InferenceClient:
-    global _hf_client
-    if _hf_client is None:
-        _hf_client = InferenceClient(provider="auto", api_key=token)
-    return _hf_client
+    resp = requests.post(url, json=payload, timeout=45)
+    resp.raise_for_status()
+    data = resp.json()
+    return data["candidates"][0]["content"]["parts"][0]["text"]
 
 # ──────────────────────────────────────────────────────────────────────────────
 # IMAGE HELPERS
@@ -476,9 +508,11 @@ def validate_result(data: dict, keys: list) -> dict | None:
 
 def analyze_one(img: Image.Image, aspect: str, token: str) -> tuple:
     """
-    Try MODELS in order for a single image.
-    Returns (result_dict, model_short_name) on success,
-            (None, joined_error_string)      on total failure.
+    Try Gemini MODELS in order for a single image.
+    Returns (result_dict, model_name) on success,
+            (None, joined_error_string) on total failure.
+    Image is encoded once and reused across all fallback attempts.
+    token = GEMINI_API_KEY environment variable value.
     """
     b64    = pil_to_b64(img)
     keys   = ASPECT_KEYS[aspect]
@@ -486,53 +520,30 @@ def analyze_one(img: Image.Image, aspect: str, token: str) -> tuple:
     errors = []
 
     for model in MODELS:
-        short = model.split("/")[-1]
         try:
-            client = _get_client(token)
-            resp = client.chat_completion(
-                model=model,
-                messages=[
-                    {
-                        "role": "system",
-                        "content": (
-                            "You are a JSON-only API for trailer inspection. "
-                            "You MUST respond with a single valid flat JSON object and absolutely "
-                            "nothing else — no explanation, no preamble, no markdown fences, "
-                            "no reasoning text, no nested objects. "
-                            "Every value must be exactly the string \"detected\" or \"missing\". "
-                            "Start your response with '{' and end with '}'."
-                        ),
-                    },
-                    {
-                        "role": "user",
-                        "content": [
-                            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}},
-                            {"type": "text", "text": prompt},
-                        ],
-                    },
-                ],
-                max_tokens=120,
-                temperature=0.05,
-            )
-            raw_content = resp.choices[0].message.content
-            print(f"[{short}][{aspect}] raw: {raw_content[:300]}")
+            raw_content = call_gemini(b64, prompt, model, token)
+            print(f"[{model}][{aspect}] raw: {raw_content[:300]}")
             data   = extract_json(raw_content, keys)
             result = validate_result(data, keys)
             if result is not None:
-                return result, short
-            errors.append(f"{short}: JSON parse failed. Raw: {raw_content[:150]}")
-        except Exception as e:
-            err = str(e)
-            if "401" in err or "403" in err:
-                errors.append(f"{short}: auth error — check HF_TOKEN ({err[:100]})")
-            elif "404" in err:
-                errors.append(f"{short}: 404 — model unavailable ({err[:100]})")
-            elif "429" in err:
-                errors.append(f"{short}: rate limited — retrying next model")
-            elif "503" in err or "502" in err:
-                errors.append(f"{short}: model loading — retrying next model")
+                return result, model
+            errors.append(f"{model}: JSON parse failed. Raw: {raw_content[:150]}")
+        except requests.HTTPError as e:
+            status = e.response.status_code if e.response is not None else "?"
+            if status == 400:
+                errors.append(f"{model}: bad request — check image or prompt ({str(e)[:120]})")
+            elif status in (401, 403):
+                errors.append(f"{model}: invalid API key — check GEMINI_API_KEY")
+            elif status == 429:
+                errors.append(f"{model}: rate limited — retrying next model")
+            elif status == 503:
+                errors.append(f"{model}: service unavailable — retrying next model")
             else:
-                errors.append(f"{short}: {err[:180]}")
+                errors.append(f"{model}: HTTP {status} — {str(e)[:150]}")
+        except requests.Timeout:
+            errors.append(f"{model}: request timed out — retrying next model")
+        except Exception as e:
+            errors.append(f"{model}: {str(e)[:180]}")
 
     return None, " | ".join(errors)
 
@@ -753,11 +764,11 @@ def root():
 
 @app.get("/health", tags=["Health"])
 def health():
-    token = os.environ.get("HF_TOKEN", "").strip()
+    token = os.environ.get("GEMINI_API_KEY", "").strip()
     return {
         "status": "ok",
-        "hf_token_set": bool(token),
-        "models": [m.split("/")[-1] for m in MODELS],
+        "gemini_api_key_set": bool(token),
+        "models": MODELS,
     }
 
 
@@ -771,13 +782,14 @@ def inspect(request: InspectRequest):
 
     Labels accepted: `front_right`, `front_left`, `rear_right`, `rear_left`, `inside`, `door`
     """
-    token = os.environ.get("HF_TOKEN", "").strip()
+    token = os.environ.get("GEMINI_API_KEY", "").strip()
     if not token:
         raise HTTPException(
             status_code=503,
             detail=(
-                "HF_TOKEN not configured. "
-                "Set it in Space Settings → Repository Secrets."
+                "GEMINI_API_KEY not configured. "
+                "Set it in Space Settings → Repository Secrets. "
+                "Get a free key at https://aistudio.google.com/apikey"
             ),
         )
 
@@ -911,11 +923,11 @@ def inspect(request: InspectRequest):
 # STARTUP
 # ──────────────────────────────────────────────────────────────────────────────
 
-_tok = os.environ.get("HF_TOKEN", "")
+_tok = os.environ.get("GEMINI_API_KEY", "")
 print("=" * 60)
-print("  Amazon Trailer Inspector — API mode")
-print(f"  HF_TOKEN : {'SET (' + str(len(_tok)) + ' chars)' if _tok else 'NOT SET ⚠️'}")
-print(f"  Models   : {[m.split('/')[-1] for m in MODELS]}")
+print("  Amazon Trailer Inspector — API mode (Gemini)")
+print(f"  GEMINI_API_KEY : {'SET (' + str(len(_tok)) + ' chars)' if _tok else 'NOT SET ⚠️  → get free key at aistudio.google.com/apikey'}")
+print(f"  Models         : {MODELS}")
 print("=" * 60)
 
 if __name__ == "__main__":