prompture 0.0.35__py3-none-any.whl → 0.0.40.dev1__py3-none-any.whl

prompture/drivers/async_openai_driver.py

@@ -2,7 +2,9 @@
 
 from __future__ import annotations
 
+import json
 import os
+from collections.abc import AsyncIterator
 from typing import Any
 
 try:
@@ -18,6 +20,9 @@ from .openai_driver import OpenAIDriver
 class AsyncOpenAIDriver(CostMixin, AsyncDriver):
     supports_json_mode = True
     supports_json_schema = True
+    supports_tool_use = True
+    supports_streaming = True
+    supports_vision = True
 
     MODEL_PRICING = OpenAIDriver.MODEL_PRICING
 
@@ -31,12 +36,17 @@ class AsyncOpenAIDriver(CostMixin, AsyncDriver):
 
     supports_messages = True
 
+    def _prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        from .vision_helpers import _prepare_openai_vision_messages
+
+        return _prepare_openai_vision_messages(messages)
+
     async def generate(self, prompt: str, options: dict[str, Any]) -> dict[str, Any]:
         messages = [{"role": "user", "content": prompt}]
         return await self._do_generate(messages, options)
 
     async def generate_messages(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
-        return await self._do_generate(messages, options)
+        return await self._do_generate(self._prepare_messages(messages), options)
 
     async def _do_generate(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
         if self.client is None:
@@ -44,9 +54,16 @@ class AsyncOpenAIDriver(CostMixin, AsyncDriver):
 
         model = options.get("model", self.model)
 
-        model_info = self.MODEL_PRICING.get(model, {})
-        tokens_param = model_info.get("tokens_param", "max_tokens")
-        supports_temperature = model_info.get("supports_temperature", True)
+        model_config = self._get_model_config("openai", model)
+        tokens_param = model_config["tokens_param"]
+        supports_temperature = model_config["supports_temperature"]
+
+        # Validate capabilities against models.dev metadata
+        self._validate_model_capabilities(
+            "openai",
+            model,
+            using_json_schema=bool(options.get("json_schema")),
+        )
 
         opts = {"temperature": 1.0, "max_tokens": 512, **options}
 
@@ -87,10 +104,150 @@ class AsyncOpenAIDriver(CostMixin, AsyncDriver):
             "prompt_tokens": prompt_tokens,
             "completion_tokens": completion_tokens,
             "total_tokens": total_tokens,
-            "cost": total_cost,
+            "cost": round(total_cost, 6),
             "raw_response": resp.model_dump(),
             "model_name": model,
         }
 
         text = resp.choices[0].message.content
         return {"text": text, "meta": meta}
+
+    # ------------------------------------------------------------------
+    # Tool use
+    # ------------------------------------------------------------------
+
+    async def generate_messages_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        options: dict[str, Any],
+    ) -> dict[str, Any]:
+        """Generate a response that may include tool calls."""
+        if self.client is None:
+            raise RuntimeError("openai package (>=1.0.0) is not installed")
+
+        model = options.get("model", self.model)
+        model_config = self._get_model_config("openai", model)
+        tokens_param = model_config["tokens_param"]
+        supports_temperature = model_config["supports_temperature"]
+
+        self._validate_model_capabilities("openai", model, using_tool_use=True)
+
+        opts = {"temperature": 1.0, "max_tokens": 512, **options}
+
+        kwargs: dict[str, Any] = {
+            "model": model,
+            "messages": messages,
+            "tools": tools,
+        }
+        kwargs[tokens_param] = opts.get("max_tokens", 512)
+
+        if supports_temperature and "temperature" in opts:
+            kwargs["temperature"] = opts["temperature"]
+
+        resp = await self.client.chat.completions.create(**kwargs)
+
+        usage = getattr(resp, "usage", None)
+        prompt_tokens = getattr(usage, "prompt_tokens", 0)
+        completion_tokens = getattr(usage, "completion_tokens", 0)
+        total_tokens = getattr(usage, "total_tokens", 0)
+        total_cost = self._calculate_cost("openai", model, prompt_tokens, completion_tokens)
+
+        meta = {
+            "prompt_tokens": prompt_tokens,
+            "completion_tokens": completion_tokens,
+            "total_tokens": total_tokens,
+            "cost": round(total_cost, 6),
+            "raw_response": resp.model_dump(),
+            "model_name": model,
+        }
+
+        choice = resp.choices[0]
+        text = choice.message.content or ""
+        stop_reason = choice.finish_reason
+
+        tool_calls_out: list[dict[str, Any]] = []
+        if choice.message.tool_calls:
+            for tc in choice.message.tool_calls:
+                try:
+                    args = json.loads(tc.function.arguments)
+                except (json.JSONDecodeError, TypeError):
+                    args = {}
+                tool_calls_out.append({
+                    "id": tc.id,
+                    "name": tc.function.name,
+                    "arguments": args,
+                })
+
+        return {
+            "text": text,
+            "meta": meta,
+            "tool_calls": tool_calls_out,
+            "stop_reason": stop_reason,
+        }
+
+    # ------------------------------------------------------------------
+    # Streaming
+    # ------------------------------------------------------------------
+
+    async def generate_messages_stream(
+        self,
+        messages: list[dict[str, Any]],
+        options: dict[str, Any],
+    ) -> AsyncIterator[dict[str, Any]]:
+        """Yield response chunks via OpenAI streaming API."""
+        if self.client is None:
+            raise RuntimeError("openai package (>=1.0.0) is not installed")
+
+        model = options.get("model", self.model)
+        model_config = self._get_model_config("openai", model)
+        tokens_param = model_config["tokens_param"]
+        supports_temperature = model_config["supports_temperature"]
+
+        opts = {"temperature": 1.0, "max_tokens": 512, **options}
+
+        kwargs: dict[str, Any] = {
+            "model": model,
+            "messages": messages,
+            "stream": True,
+            "stream_options": {"include_usage": True},
+        }
+        kwargs[tokens_param] = opts.get("max_tokens", 512)
+
+        if supports_temperature and "temperature" in opts:
+            kwargs["temperature"] = opts["temperature"]
+
+        stream = await self.client.chat.completions.create(**kwargs)
+
+        full_text = ""
+        prompt_tokens = 0
+        completion_tokens = 0
+
+        async for chunk in stream:
+            # Usage comes in the final chunk
+            if getattr(chunk, "usage", None):
+                prompt_tokens = chunk.usage.prompt_tokens or 0
+                completion_tokens = chunk.usage.completion_tokens or 0
+
+            if chunk.choices:
+                delta = chunk.choices[0].delta
+                content = getattr(delta, "content", None) or ""
+                if content:
+                    full_text += content
+                    yield {"type": "delta", "text": content}
+
+        total_tokens = prompt_tokens + completion_tokens
+        total_cost = self._calculate_cost("openai", model, prompt_tokens, completion_tokens)
+
+        yield {
+            "type": "done",
+            "text": full_text,
+            "meta": {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": total_tokens,
+                "cost": round(total_cost, 6),
+                "raw_response": {},
+                "model_name": model,
+            },
+        }

prompture/drivers/async_openrouter_driver.py

@@ -14,6 +14,7 @@ from .openrouter_driver import OpenRouterDriver
 
 class AsyncOpenRouterDriver(CostMixin, AsyncDriver):
     supports_json_mode = True
+    supports_vision = True
 
     MODEL_PRICING = OpenRouterDriver.MODEL_PRICING
 
@@ -31,19 +32,24 @@ class AsyncOpenRouterDriver(CostMixin, AsyncDriver):
 
     supports_messages = True
 
+    def _prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        from .vision_helpers import _prepare_openai_vision_messages
+
+        return _prepare_openai_vision_messages(messages)
+
     async def generate(self, prompt: str, options: dict[str, Any]) -> dict[str, Any]:
         messages = [{"role": "user", "content": prompt}]
         return await self._do_generate(messages, options)
 
     async def generate_messages(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
-        return await self._do_generate(messages, options)
+        return await self._do_generate(self._prepare_messages(messages), options)
 
     async def _do_generate(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
         model = options.get("model", self.model)
 
-        model_info = self.MODEL_PRICING.get(model, {})
-        tokens_param = model_info.get("tokens_param", "max_tokens")
-        supports_temperature = model_info.get("supports_temperature", True)
+        model_config = self._get_model_config("openrouter", model)
+        tokens_param = model_config["tokens_param"]
+        supports_temperature = model_config["supports_temperature"]
 
         opts = {"temperature": 1.0, "max_tokens": 512, **options}
 
@@ -87,7 +93,7 @@ class AsyncOpenRouterDriver(CostMixin, AsyncDriver):
             "prompt_tokens": prompt_tokens,
             "completion_tokens": completion_tokens,
             "total_tokens": total_tokens,
-            "cost": total_cost,
+            "cost": round(total_cost, 6),
             "raw_response": resp,
             "model_name": model,
         }

prompture/drivers/async_registry.py

@@ -49,7 +49,11 @@ register_async_driver(
 )
 register_async_driver(
     "lmstudio",
-    lambda model=None: AsyncLMStudioDriver(endpoint=settings.lmstudio_endpoint, model=model or settings.lmstudio_model),
+    lambda model=None: AsyncLMStudioDriver(
+        endpoint=settings.lmstudio_endpoint,
+        model=model or settings.lmstudio_model,
+        api_key=settings.lmstudio_api_key,
+    ),
     overwrite=True,
 )
 register_async_driver(

prompture/drivers/azure_driver.py

@@ -17,6 +17,7 @@ from ..driver import Driver
 class AzureDriver(CostMixin, Driver):
     supports_json_mode = True
     supports_json_schema = True
+    supports_vision = True
 
     # Pricing per 1K tokens (adjust if your Azure pricing differs from OpenAI defaults)
     MODEL_PRICING = {
@@ -90,21 +91,26 @@ class AzureDriver(CostMixin, Driver):
 
     supports_messages = True
 
+    def _prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        from .vision_helpers import _prepare_openai_vision_messages
+
+        return _prepare_openai_vision_messages(messages)
+
     def generate(self, prompt: str, options: dict[str, Any]) -> dict[str, Any]:
         messages = [{"role": "user", "content": prompt}]
         return self._do_generate(messages, options)
 
     def generate_messages(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
-        return self._do_generate(messages, options)
+        return self._do_generate(self._prepare_messages(messages), options)
 
     def _do_generate(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
         if self.client is None:
             raise RuntimeError("openai package (>=1.0.0) with AzureOpenAI not installed")
 
         model = options.get("model", self.model)
-        model_info = self.MODEL_PRICING.get(model, {})
-        tokens_param = model_info.get("tokens_param", "max_tokens")
-        supports_temperature = model_info.get("supports_temperature", True)
+        model_config = self._get_model_config("azure", model)
+        tokens_param = model_config["tokens_param"]
+        supports_temperature = model_config["supports_temperature"]
 
         opts = {"temperature": 1.0, "max_tokens": 512, **options}
 

prompture/drivers/claude_driver.py

@@ -21,6 +21,7 @@ class ClaudeDriver(CostMixin, Driver):
     supports_json_schema = True
     supports_tool_use = True
     supports_streaming = True
+    supports_vision = True
 
     # Claude pricing per 1000 tokens (prices should be kept current with Anthropic's pricing)
     MODEL_PRICING = {
@@ -57,12 +58,17 @@ class ClaudeDriver(CostMixin, Driver):
 
     supports_messages = True
 
+    def _prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        from .vision_helpers import _prepare_claude_vision_messages
+
+        return _prepare_claude_vision_messages(messages)
+
     def generate(self, prompt: str, options: dict[str, Any]) -> dict[str, Any]:
         messages = [{"role": "user", "content": prompt}]
         return self._do_generate(messages, options)
 
     def generate_messages(self, messages: list[dict[str, Any]], options: dict[str, Any]) -> dict[str, Any]:
-        return self._do_generate(messages, options)
+        return self._do_generate(self._prepare_messages(messages), options)
 
     def _do_generate(self, messages: list[dict[str, Any]], options: dict[str, Any]) -> dict[str, Any]:
         if anthropic is None:
@@ -71,6 +77,13 @@ class ClaudeDriver(CostMixin, Driver):
         opts = {**{"temperature": 0.0, "max_tokens": 512}, **options}
         model = options.get("model", self.model)
 
+        # Validate capabilities against models.dev metadata
+        self._validate_model_capabilities(
+            "claude",
+            model,
+            using_json_schema=bool(options.get("json_schema")),
+        )
+
         client = anthropic.Anthropic(api_key=self.api_key)
 
         # Anthropic requires system messages as a top-level parameter
@@ -171,6 +184,9 @@ class ClaudeDriver(CostMixin, Driver):
 
         opts = {**{"temperature": 0.0, "max_tokens": 512}, **options}
         model = options.get("model", self.model)
+
+        self._validate_model_capabilities("claude", model, using_tool_use=True)
+
         client = anthropic.Anthropic(api_key=self.api_key)
 
         system_content, api_messages = self._extract_system_and_messages(messages)

prompture/drivers/google_driver.py

@@ -1,5 +1,7 @@
 import logging
 import os
+import uuid
+from collections.abc import Iterator
 from typing import Any, Optional
 
 import google.generativeai as genai
@@ -15,6 +17,9 @@ class GoogleDriver(CostMixin, Driver):
 
     supports_json_mode = True
     supports_json_schema = True
+    supports_vision = True
+    supports_tool_use = True
+    supports_streaming = True
 
     # Based on current Gemini pricing (as of 2025)
     # Source: https://cloud.google.com/vertex-ai/pricing#gemini_models
@@ -105,25 +110,62 @@ class GoogleDriver(CostMixin, Driver):
             completion_cost = (completion_chars / 1_000_000) * model_pricing["completion"]
         return round(prompt_cost + completion_cost, 6)
 
+    def _extract_usage_metadata(self, response: Any, messages: list[dict[str, Any]]) -> dict[str, Any]:
+        """Extract token counts from response, falling back to character estimation."""
+        usage = getattr(response, "usage_metadata", None)
+        if usage:
+            prompt_tokens = getattr(usage, "prompt_token_count", 0) or 0
+            completion_tokens = getattr(usage, "candidates_token_count", 0) or 0
+            total_tokens = getattr(usage, "total_token_count", 0) or (prompt_tokens + completion_tokens)
+            cost = self._calculate_cost("google", self.model, prompt_tokens, completion_tokens)
+        else:
+            # Fallback: estimate from character counts
+            total_prompt_chars = 0
+            for msg in messages:
+                c = msg.get("content", "")
+                if isinstance(c, str):
+                    total_prompt_chars += len(c)
+                elif isinstance(c, list):
+                    for part in c:
+                        if isinstance(part, str):
+                            total_prompt_chars += len(part)
+                        elif isinstance(part, dict) and "text" in part:
+                            total_prompt_chars += len(part["text"])
+            completion_chars = len(response.text) if response.text else 0
+            prompt_tokens = total_prompt_chars // 4
+            completion_tokens = completion_chars // 4
+            total_tokens = prompt_tokens + completion_tokens
+            cost = self._calculate_cost_chars(total_prompt_chars, completion_chars)
+
+        return {
+            "prompt_tokens": prompt_tokens,
+            "completion_tokens": completion_tokens,
+            "total_tokens": total_tokens,
+            "cost": round(cost, 6),
+        }
+
     supports_messages = True
 
-    def generate(self, prompt: str, options: Optional[dict[str, Any]] = None) -> dict[str, Any]:
-        messages = [{"role": "user", "content": prompt}]
-        return self._do_generate(messages, options)
+    def _prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        from .vision_helpers import _prepare_google_vision_messages
 
-    def generate_messages(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
-        return self._do_generate(messages, options)
+        return _prepare_google_vision_messages(messages)
 
-    def _do_generate(self, messages: list[dict[str, str]], options: Optional[dict[str, Any]] = None) -> dict[str, Any]:
+    def _build_generation_args(
+        self, messages: list[dict[str, Any]], options: Optional[dict[str, Any]] = None
+    ) -> tuple[Any, dict[str, Any]]:
+        """Parse messages and options into (gen_input, kwargs) for generate_content.
+
+        Returns the content input and a dict of keyword arguments
+        (generation_config, safety_settings, model kwargs including system_instruction).
+        """
         merged_options = self.options.copy()
         if options:
             merged_options.update(options)
 
-        # Extract specific options for Google's API
         generation_config = merged_options.get("generation_config", {})
         safety_settings = merged_options.get("safety_settings", {})
 
-        # Map common options to generation_config if not present
         if "temperature" in merged_options and "temperature" not in generation_config:
             generation_config["temperature"] = merged_options["temperature"]
         if "max_tokens" in merged_options and "max_output_tokens" not in generation_config:
@@ -147,44 +189,66 @@ class GoogleDriver(CostMixin, Driver):
             role = msg.get("role", "user")
             content = msg.get("content", "")
             if role == "system":
-                system_instruction = content
+                system_instruction = content if isinstance(content, str) else str(content)
             else:
-                # Gemini uses "model" for assistant role
                 gemini_role = "model" if role == "assistant" else "user"
-                contents.append({"role": gemini_role, "parts": [content]})
+                if msg.get("_vision_parts"):
+                    contents.append({"role": gemini_role, "parts": content})
+                else:
+                    contents.append({"role": gemini_role, "parts": [content]})
+
+        # For a single message, unwrap only if it has exactly one string part
+        if len(contents) == 1:
+            parts = contents[0]["parts"]
+            if len(parts) == 1 and isinstance(parts[0], str):
+                gen_input = parts[0]
+            else:
+                gen_input = contents
+        else:
+            gen_input = contents
+
+        model_kwargs: dict[str, Any] = {}
+        if system_instruction:
+            model_kwargs["system_instruction"] = system_instruction
+
+        gen_kwargs: dict[str, Any] = {
+            "generation_config": generation_config if generation_config else None,
+            "safety_settings": safety_settings if safety_settings else None,
+        }
+
+        return gen_input, gen_kwargs, model_kwargs
+
+    def generate(self, prompt: str, options: Optional[dict[str, Any]] = None) -> dict[str, Any]:
+        messages = [{"role": "user", "content": prompt}]
+        return self._do_generate(messages, options)
+
+    def generate_messages(self, messages: list[dict[str, str]], options: dict[str, Any]) -> dict[str, Any]:
+        return self._do_generate(self._prepare_messages(messages), options)
+
+    def _do_generate(self, messages: list[dict[str, str]], options: Optional[dict[str, Any]] = None) -> dict[str, Any]:
+        gen_input, gen_kwargs, model_kwargs = self._build_generation_args(messages, options)
+
+        # Validate capabilities against models.dev metadata
+        self._validate_model_capabilities(
+            "google",
+            self.model,
+            using_json_schema=bool((options or {}).get("json_schema")),
+        )
 
         try:
             logger.debug(f"Initializing {self.model} for generation")
-            model_kwargs: dict[str, Any] = {}
-            if system_instruction:
-                model_kwargs["system_instruction"] = system_instruction
             model = genai.GenerativeModel(self.model, **model_kwargs)
 
-            # Generate response
-            logger.debug(f"Generating with {len(contents)} content parts")
-            # If single user message, pass content directly for backward compatibility
-            gen_input: Any = contents if len(contents) != 1 else contents[0]["parts"][0]
-            response = model.generate_content(
-                gen_input,
-                generation_config=generation_config if generation_config else None,
-                safety_settings=safety_settings if safety_settings else None,
-            )
+            logger.debug(f"Generating with model {self.model}")
+            response = model.generate_content(gen_input, **gen_kwargs)
 
             if not response.text:
                 raise ValueError("Empty response from model")
 
-            # Calculate token usage and cost
-            total_prompt_chars = sum(len(msg.get("content", "")) for msg in messages)
-            completion_chars = len(response.text)
-
-            # Google uses character-based cost estimation
-            total_cost = self._calculate_cost_chars(total_prompt_chars, completion_chars)
+            usage_meta = self._extract_usage_metadata(response, messages)
 
             meta = {
-                "prompt_chars": total_prompt_chars,
-                "completion_chars": completion_chars,
-                "total_chars": total_prompt_chars + completion_chars,
-                "cost": total_cost,
+                **usage_meta,
                 "raw_response": response.prompt_feedback if hasattr(response, "prompt_feedback") else None,
                 "model_name": self.model,
             }
@@ -194,3 +258,133 @@ class GoogleDriver(CostMixin, Driver):
         except Exception as e:
             logger.error(f"Google API request failed: {e}")
             raise RuntimeError(f"Google API request failed: {e}") from e
+
+    # ------------------------------------------------------------------
+    # Tool use
+    # ------------------------------------------------------------------
+
+    def generate_messages_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        options: dict[str, Any],
+    ) -> dict[str, Any]:
+        """Generate a response that may include tool/function calls."""
+        model = options.get("model", self.model)
+        self._validate_model_capabilities("google", model, using_tool_use=True)
+
+        gen_input, gen_kwargs, model_kwargs = self._build_generation_args(
+            self._prepare_messages(messages), options
+        )
+
+        # Convert tools from OpenAI format to Gemini function declarations
+        function_declarations = []
+        for t in tools:
+            if "type" in t and t["type"] == "function":
+                fn = t["function"]
+                decl = {
+                    "name": fn["name"],
+                    "description": fn.get("description", ""),
+                }
+                params = fn.get("parameters")
+                if params:
+                    decl["parameters"] = params
+                function_declarations.append(decl)
+            elif "name" in t:
+                # Already in a generic format
+                decl = {"name": t["name"], "description": t.get("description", "")}
+                params = t.get("parameters") or t.get("input_schema")
+                if params:
+                    decl["parameters"] = params
+                function_declarations.append(decl)
+
+        try:
+            model = genai.GenerativeModel(self.model, **model_kwargs)
+
+            gemini_tools = [genai.types.Tool(function_declarations=function_declarations)]
+            response = model.generate_content(gen_input, tools=gemini_tools, **gen_kwargs)
+
+            usage_meta = self._extract_usage_metadata(response, messages)
+            meta = {
+                **usage_meta,
+                "raw_response": response.prompt_feedback if hasattr(response, "prompt_feedback") else None,
+                "model_name": self.model,
+            }
+
+            text = ""
+            tool_calls_out: list[dict[str, Any]] = []
+            stop_reason = "stop"
+
+            for candidate in response.candidates:
+                for part in candidate.content.parts:
+                    if hasattr(part, "text") and part.text:
+                        text += part.text
+                    if hasattr(part, "function_call") and part.function_call.name:
+                        fc = part.function_call
+                        tool_calls_out.append({
+                            "id": str(uuid.uuid4()),
+                            "name": fc.name,
+                            "arguments": dict(fc.args) if fc.args else {},
+                        })
+
+                finish_reason = getattr(candidate, "finish_reason", None)
+                if finish_reason is not None:
+                    # Map Gemini finish reasons to standard stop reasons
+                    reason_map = {1: "stop", 2: "max_tokens", 3: "safety", 4: "recitation", 5: "other"}
+                    stop_reason = reason_map.get(finish_reason, "stop")
+
+            if tool_calls_out:
+                stop_reason = "tool_use"
+
+            return {
+                "text": text,
+                "meta": meta,
+                "tool_calls": tool_calls_out,
+                "stop_reason": stop_reason,
+            }
+
+        except Exception as e:
+            logger.error(f"Google API tool call request failed: {e}")
+            raise RuntimeError(f"Google API tool call request failed: {e}") from e
+
+    # ------------------------------------------------------------------
+    # Streaming
+    # ------------------------------------------------------------------
+
+    def generate_messages_stream(
+        self,
+        messages: list[dict[str, Any]],
+        options: dict[str, Any],
+    ) -> Iterator[dict[str, Any]]:
+        """Yield response chunks via Gemini streaming API."""
+        gen_input, gen_kwargs, model_kwargs = self._build_generation_args(
+            self._prepare_messages(messages), options
+        )
+
+        try:
+            model = genai.GenerativeModel(self.model, **model_kwargs)
+            response = model.generate_content(gen_input, stream=True, **gen_kwargs)
+
+            full_text = ""
+            for chunk in response:
+                chunk_text = getattr(chunk, "text", None) or ""
+                if chunk_text:
+                    full_text += chunk_text
+                    yield {"type": "delta", "text": chunk_text}
+
+            # After iteration completes, resolve() has been called on the response
+            usage_meta = self._extract_usage_metadata(response, messages)
+
+            yield {
+                "type": "done",
+                "text": full_text,
+                "meta": {
+                    **usage_meta,
+                    "raw_response": {},
+                    "model_name": self.model,
+                },
+            }
+
+        except Exception as e:
+            logger.error(f"Google API streaming request failed: {e}")
+            raise RuntimeError(f"Google API streaming request failed: {e}") from e