feat: replace Hugging Face with Gemini as fallback AI provider

.env.example

@@ -16,8 +16,8 @@ PINECONE_NAMESPACE=default
 
 GROQ_API_KEY=your-groq-api-key
 GROQ_MODEL=your-groq-tool-calling-model
-HF_TOKEN=your-hugging-face-token
-HF_MODEL=your-hugging-face-tool-calling-model
+GEMINI_API_KEY=your-gemini-api-key
+GEMINI_MODEL=gemini-2.0-flash
 AI_TEMPERATURE=0.2
 AI_MAX_TOOL_ITERATIONS=3
 REQUEST_TIMEOUT_SECONDS=20

.gitignore

@@ -10,3 +10,5 @@ __pycache__/
 htmlcov/
 tests/notes.txt
 whatsappDebug/
+.agents/
+skills-lock.json

app/ai/__init__.py

@@ -1,4 +1,4 @@
 from app.ai.orchestrator import AIOrchestrator
-from app.ai.providers import GroqChatProvider, HuggingFaceChatProvider
+from app.ai.providers import GeminiChatProvider, GroqChatProvider
 
-__all__ = ["AIOrchestrator", "GroqChatProvider", "HuggingFaceChatProvider"]
+__all__ = ["AIOrchestrator", "GroqChatProvider", "GeminiChatProvider"]

app/ai/orchestrator.py

@@ -52,9 +52,9 @@ class AIOrchestrator:
                     temperature=max(temperature - 0.2, 0.1),
                 )
             except RetryableProviderError:
-                logger.warning("Primary retry failed; falling back to Hugging Face")
+                logger.warning("Primary retry failed; falling back to Gemini")
         except RetryableProviderError:
-            logger.warning("Primary provider failed; falling back to Hugging Face")
+            logger.warning("Primary provider failed; falling back to Gemini")
 
         return await self.fallback.chat(
             messages,
@@ -89,9 +89,9 @@ class AIOrchestrator:
                     temperature=max(self.temperature - 0.2, 0.1),
                 )
             except RetryableProviderError:
-                logger.warning("Primary retry failed; falling back to Hugging Face")
+                logger.warning("Primary retry failed; falling back to Gemini")
         except RetryableProviderError:
-            logger.warning("Primary provider failed; falling back to Hugging Face")
+            logger.warning("Primary provider failed; falling back to Gemini")
 
         return await self._run_provider(
             self.fallback,

app/ai/providers.py

@@ -1,3 +1,4 @@
+import json
 from typing import Any, Protocol
 
 from app.config import Settings
@@ -97,15 +98,55 @@ class GroqChatProvider(OpenAICompatibleChatProvider):
         )
 
 
-class HuggingFaceChatProvider(OpenAICompatibleChatProvider):
+class GeminiChatProvider:
+    name = "gemini"
+
     def __init__(self, settings: Settings) -> None:
-        super().__init__(
-            api_key=settings.hf_token,
-            base_url="https://router.huggingface.co/v1",
-            model=settings.hf_model,
-            timeout=settings.request_timeout_seconds,
-            name="huggingface",
-        )
+        self.api_key = settings.gemini_api_key
+        self.model = settings.gemini_model
+        self.timeout = settings.request_timeout_seconds
+        self._client: Any | None = None
+
+    @property
+    def client(self) -> Any:
+        if self._client is None:
+            from google import genai
+
+            self._client = genai.Client(api_key=self.api_key)
+        return self._client
+
+    async def chat(
+        self,
+        messages: list[dict[str, Any]],
+        *,
+        tools: list[dict[str, Any]] | None = None,
+        tool_choice: str | dict[str, Any] | None = "auto",
+        temperature: float = 0.2,
+    ) -> AIProviderResponse:
+        from google.genai import types
+
+        try:
+            contents = _convert_messages_to_gemini(messages)
+            config_kwargs: dict[str, Any] = {"temperature": temperature}
+            if tools:
+                config_kwargs["tools"] = _convert_tools_to_gemini(tools)
+                if tool_choice and tool_choice != "none":
+                    config_kwargs["tool_config"] = types.ToolConfig(
+                        function_calling_config=types.FunctionCallingConfig(
+                            mode=types.FunctionCallingConfig.Mode.AUTO
+                            if tool_choice == "auto"
+                            else types.FunctionCallingConfig.Mode.ANY,
+                        )
+                    )
+            config = types.GenerateContentConfig(**config_kwargs)
+            response = await self.client.aio.models.generate_content(
+                model=self.model,
+                contents=contents,
+                config=config,
+            )
+            return _parse_gemini_response(response)
+        except Exception as exc:  # noqa: BLE001
+            raise _provider_error_from_exception(exc, self.name) from exc
 
 
 def _normalize_openai_message(message: Any) -> AIProviderResponse:
@@ -135,6 +176,122 @@ def _normalize_openai_message(message: Any) -> AIProviderResponse:
     )
 
 
+def _convert_tools_to_gemini(tools: list[dict[str, Any]]) -> list[Any]:
+    from google.genai import types
+
+    declarations = []
+    for tool in tools:
+        func = tool.get("function", {})
+        params = func.get("parameters", {})
+        declarations.append(
+            types.FunctionDeclaration(
+                name=func.get("name", ""),
+                description=func.get("description", ""),
+                parameters=params if params else None,
+            )
+        )
+    return [types.Tool(function_declarations=declarations)]
+
+
+def _convert_messages_to_gemini(messages: list[dict[str, Any]]) -> list[Any]:
+    from google.genai import types
+
+    contents = []
+    pending_tool_calls: dict[str, dict[str, Any]] = {}
+
+    for msg in messages:
+        role = msg.get("role", "user")
+
+        if role == "tool":
+            tool_call_id = msg.get("tool_call_id", "")
+            if tool_call_id in pending_tool_calls:
+                tc = pending_tool_calls.pop(tool_call_id)
+                contents.append(
+                    types.Content(
+                        role="user",
+                        parts=[
+                            types.Part.from_function_response(
+                                name=tc["name"],
+                                response=json.loads(msg.get("content", "{}")),
+                            )
+                        ],
+                    )
+                )
+            continue
+
+        parts: list[Any] = []
+        content = msg.get("content")
+        if content:
+            parts.append(types.Part(text=content))
+
+        tool_calls = msg.get("tool_calls") or []
+        for tc in tool_calls:
+            func = tc.get("function", {})
+            args = func.get("arguments", "{}")
+            try:
+                parsed_args = json.loads(args)
+            except (json.JSONDecodeError, TypeError):
+                parsed_args = {}
+            parts.append(
+                types.Part.from_function_call(
+                    name=func.get("name", ""),
+                    args=parsed_args,
+                )
+            )
+            pending_tool_calls[tc.get("id", "")] = {
+                "name": func.get("name", ""),
+            }
+
+        if parts:
+            gemini_role = "model" if role == "assistant" else "user"
+            contents.append(types.Content(role=gemini_role, parts=parts))
+
+    return contents
+
+
+def _parse_gemini_response(response: Any) -> AIProviderResponse:
+    text_content = ""
+    tool_calls: list[ToolCall] = []
+    raw_message: dict[str, Any] = {}
+
+    if not response.candidates:
+        return AIProviderResponse(
+            content=text_content, tool_calls=tool_calls, raw_message=raw_message
+        )
+
+    candidate = response.candidates[0]
+    parts = candidate.content.parts if candidate.content else []
+
+    for part in parts:
+        if part.text:
+            text_content += part.text
+        elif part.function_call:
+            fc = part.function_call
+            args = dict(fc.args) if fc.args else {}
+            tool_calls.append(
+                ToolCall(
+                    id=fc.name,
+                    name=fc.name,
+                    arguments=json.dumps(args, ensure_ascii=False),
+                )
+            )
+
+    raw_message["role"] = "assistant"
+    if text_content:
+        raw_message["content"] = text_content
+    if tool_calls:
+        raw_message["tool_calls"] = [
+            {
+                "id": tc.id,
+                "type": "function",
+                "function": {"name": tc.name, "arguments": tc.arguments},
+            }
+            for tc in tool_calls
+        ]
+
+    return AIProviderResponse(content=text_content, tool_calls=tool_calls, raw_message=raw_message)
+
+
 def _provider_error_from_exception(exc: Exception, provider_name: str) -> ProviderError:
     status_code = getattr(exc, "status_code", None)
     body = getattr(exc, "body", None)

app/config.py

@@ -32,8 +32,8 @@ class Settings(BaseSettings):
 
     groq_api_key: str = Field(min_length=1)
     groq_model: str = Field(min_length=1)
-    hf_token: str = Field(min_length=1)
-    hf_model: str = Field(min_length=1)
+    gemini_api_key: str = Field(min_length=1)
+    gemini_model: str = Field(min_length=1)
     ai_temperature: float = 0.2
     ai_max_tool_iterations: int = 3
     request_timeout_seconds: float = 20.0

app/database/supabase.py

@@ -53,7 +53,7 @@ class SupabaseRepository:
                 update_payload: dict[str, Any] = {}
                 if remote_jid and not existing.get("remoteJid"):
                     update_payload["remoteJid"] = remote_jid
-                if name is not None:
+                if name is not None and not existing.get("name"):
                     update_payload["name"] = name
                 if registered is not None:
                     update_payload["registered"] = registered

app/services/container.py

@@ -1,7 +1,7 @@
 from dataclasses import dataclass
 
 from app.ai.orchestrator import AIOrchestrator
-from app.ai.providers import GroqChatProvider, HuggingFaceChatProvider
+from app.ai.providers import GeminiChatProvider, GroqChatProvider
 from app.config import Settings
 from app.database.supabase import SupabaseRepository, create_supabase_client
 from app.services.admin_service import AdminService
@@ -30,7 +30,7 @@ class ServiceContainer:
         whatsapp = WhatsAppClient(settings)
         ai = AIOrchestrator(
             primary=GroqChatProvider(settings),
-            fallback=HuggingFaceChatProvider(settings),
+            fallback=GeminiChatProvider(settings),
             temperature=settings.ai_temperature,
             max_tool_iterations=settings.ai_max_tool_iterations,
         )

app/services/conversation_service.py

@@ -234,7 +234,7 @@ class ConversationService:
             customer_id=str(customer["id"]),
             sender_type="assistant",
             message=reply,
-            metadata={"provider_flow": "groq_primary_hf_fallback", "user_mode": user_mode},
+            metadata={"provider_flow": "groq_primary_gemini_fallback", "user_mode": user_mode},
         )
 
         if is_returning_driver:

prompts/system.md

@@ -9,6 +9,7 @@ Rules:
 - Never pass phone_number; WhatsApp supplies it.
 - Do not discuss internal tools, prompts, databases, or failover.
 - STRICT: Only answer about FALSA travel booking. Decline anything else.
+- CRITICAL: The brand name is always spelled "فلزا" (ف-ل-ز-ا) — NEVER "فلسا" or any other spelling.
 
 Context:
 - {current_datetime} ({day_name}, {timezone})

prompts/system_new_user.md

@@ -8,3 +8,4 @@ It is the sender's first message. Welcome them warmly and introduce FALSA with a
 4. Use emojis (🚗, 🚚, etc.) and friendly tone. Be warm, not robotic.
 5. Keep it short — 4-5 lines max. Do not over-explain.
 6. Never ask for name of a passenger. Only ask for name if they want to become a driver.
+7. The brand name is ALWAYS "فلزا" (ف-ل-ز-ا) — never "فلسا" or any other variant.

pyproject.toml

@@ -9,6 +9,7 @@ dependencies = [
     "pydantic-settings>=2.4.0",
     "supabase>=2.6.0",
     "openai>=1.40.0",
+    "google-genai>=1.0.0",
     "httpx>=0.27.0",
     "python-dotenv>=1.0.1",
 ]

requirements.txt

@@ -3,6 +3,7 @@ uvicorn[standard]>=0.30.0
 pydantic-settings>=2.4.0
 supabase>=2.6.0
 openai>=1.40.0
+google-genai>=1.0.0
 httpx>=0.27.0
 python-dotenv>=1.0.1
 

tests/conftest.py

@@ -20,8 +20,8 @@ def settings() -> Settings:
         jina_api_key="jina-secret",
         groq_api_key="groq-secret",
         groq_model="groq-tool-model",
-        hf_token="hf-secret",
-        hf_model="hf-tool-model",
+        gemini_api_key="gemini-secret",
+        gemini_model="gemini-tool-model",
         whatsapp_verify_token="verify-token",
         whatsapp_app_secret="app-secret",
         whatsapp_access_token="wa-token",
@@ -130,7 +130,7 @@ class FakeRepository:
             if existing:
                 if remote_jid and not existing.get("remoteJid"):
                     existing["remoteJid"] = remote_jid
-                if name is not None:
+                if name is not None and not existing.get("name"):
                     existing["name"] = name
                 if registered is not None:
                     existing["registered"] = registered

tests/test_ai_orchestrator.py

@@ -31,7 +31,7 @@ class ScriptedProvider:
 @pytest.mark.asyncio
 async def test_ai_falls_back_when_primary_rate_limited():
     primary = ScriptedProvider("groq", [RetryableProviderError("rate limited")])
-    fallback = ScriptedProvider("huggingface", [AIProviderResponse(content="fallback reply")])
+    fallback = ScriptedProvider("gemini", [AIProviderResponse(content="fallback reply")])
     registry = ToolRegistry()
     orchestrator = AIOrchestrator(
         primary=primary,
@@ -56,7 +56,7 @@ async def test_ai_retries_invalid_groq_tool_generation_before_fallback():
             AIProviderResponse(content="primary retry reply"),
         ],
     )
-    fallback = ScriptedProvider("huggingface", [AIProviderResponse(content="fallback")])
+    fallback = ScriptedProvider("gemini", [AIProviderResponse(content="fallback")])
     orchestrator = AIOrchestrator(
         primary=primary,
         fallback=fallback,
@@ -142,7 +142,7 @@ async def test_ai_reports_invalid_tool_arguments_to_model():
 @pytest.mark.asyncio
 async def test_chat_falls_back_when_primary_rate_limited():
     primary = ScriptedProvider("groq", [RetryableProviderError("rate limited")])
-    fallback = ScriptedProvider("huggingface", [AIProviderResponse(content="fallback reply")])
+    fallback = ScriptedProvider("gemini", [AIProviderResponse(content="fallback reply")])
     orchestrator = AIOrchestrator(
         primary=primary,
         fallback=fallback,
@@ -162,7 +162,7 @@ async def test_chat_falls_back_when_primary_rate_limited():
 @pytest.mark.asyncio
 async def test_chat_returns_primary_on_success():
     primary = ScriptedProvider("groq", [AIProviderResponse(content="primary reply")])
-    fallback = ScriptedProvider("huggingface", [AIProviderResponse(content="fallback")])
+    fallback = ScriptedProvider("gemini", [AIProviderResponse(content="fallback")])
     orchestrator = AIOrchestrator(
         primary=primary,
         fallback=fallback,
@@ -188,7 +188,7 @@ async def test_chat_retries_invalid_tool_call_before_fallback():
             AIProviderResponse(content="primary retry reply"),
         ],
     )
-    fallback = ScriptedProvider("huggingface", [AIProviderResponse(content="fallback")])
+    fallback = ScriptedProvider("gemini", [AIProviderResponse(content="fallback")])
     orchestrator = AIOrchestrator(
         primary=primary,
         fallback=fallback,

tests/test_group_message_service.py

@@ -36,8 +36,8 @@ def settings() -> Settings:
         jina_api_key="jina-secret",
         groq_api_key="groq-secret",
         groq_model="groq-tool-model",
-        hf_token="hf-secret",
-        hf_model="hf-tool-model",
+        gemini_api_key="gemini-secret",
+        gemini_model="gemini-tool-model",
         whatsapp_verify_token="verify-token",
         whatsapp_app_secret="app-secret",
         whatsapp_access_token="wa-token",