synth-ai 0.2.9.dev11__py3-none-any.whl → 0.4.1__py3-none-any.whl

synth_ai/v0/lm/vendors/openai_standard_responses.py

@@ -1,259 +0,0 @@
-"""
-OpenAI Responses API extensions for OpenAIStandard vendor.
-
-This module contains the Responses API and Harmony encoding methods
-that extend the OpenAIStandard class functionality.
-"""
-
-import uuid
-from typing import Any
-
-from synth_ai.v0.lm.tools.base import BaseTool
-from synth_ai.v0.lm.vendors.base import BaseLMResponse
-
-
-def _silent_backoff_handler(_details):
-    """No-op handler to keep stdout clean while still allowing visibility via logging if desired."""
-    pass
-
-
-DEFAULT_EXCEPTIONS_TO_RETRY = (
-    Exception,  # Will be more specific when imported
-)
-
-
-class OpenAIResponsesAPIMixin:
-    """Mixin class providing Responses API functionality for OpenAI vendors."""
-
-    async def _hit_api_async_responses(
-        self,
-        model: str,
-        messages: list[dict[str, Any]],
-        lm_config: dict[str, Any],
-        previous_response_id: str | None = None,
-        use_ephemeral_cache_only: bool = False,
-        tools: list[BaseTool] | None = None,
-    ) -> BaseLMResponse:
-        """Use OpenAI Responses API for supported models."""
-
-        print(f"🔍 RESPONSES API: Called for model {model}")
-        print(f"🔍 RESPONSES API: previous_response_id = {previous_response_id}")
-
-        # Check if the client has responses attribute
-        if not hasattr(self.async_client, "responses"):
-            print("🔍 RESPONSES API: Client doesn't have responses attribute, using fallback")
-            # Fallback - use chat completions with simulated response_id
-            response = await self._hit_api_async(
-                model=model,
-                messages=messages,
-                lm_config=lm_config,
-                use_ephemeral_cache_only=use_ephemeral_cache_only,
-                tools=tools,
-            )
-
-            # Add Responses API fields
-            if not response.response_id:
-                import uuid
-
-                response.response_id = str(uuid.uuid4())
-            response.api_type = "responses"
-            return response
-
-        # Use the official Responses API
-        try:
-            # Common API call params for Responses API
-            api_params = {
-                "model": model,
-            }
-
-            # For Responses API, we use 'input' parameter
-            if previous_response_id:
-                # Continue existing thread
-                api_params["previous_response_id"] = previous_response_id
-                # Only pass the new user input
-                if messages and len(messages) > 0:
-                    # Get the last user message content
-                    last_message = messages[-1]
-                    api_params["input"] = last_message.get("content", "")
-            else:
-                # Start new thread - combine system and user messages into input
-                if messages and len(messages) > 0:
-                    # Combine messages into a single input string
-                    input_parts = []
-                    for msg in messages:
-                        role = msg.get("role", "")
-                        content = msg.get("content", "")
-                        if role == "system":
-                            input_parts.append(f"System: {content}")
-                        elif role == "user":
-                            input_parts.append(f"User: {content}")
-                        elif role == "assistant":
-                            input_parts.append(f"Assistant: {content}")
-                    api_params["input"] = "\n".join(input_parts)
-
-            # Add tools if provided
-            if tools and all(isinstance(tool, BaseTool) for tool in tools):
-                api_params["tools"] = [tool.to_openai_tool() for tool in tools]
-            elif tools:
-                api_params["tools"] = tools
-
-            # Add other parameters from lm_config if needed
-            if "max_tokens" in lm_config:
-                api_params["max_tokens"] = lm_config["max_tokens"]
-
-            print(f"🔍 RESPONSES API: Calling with params: {list(api_params.keys())}")
-
-            # Call the Responses API
-            response = await self.async_client.responses.create(**api_params)
-
-            print(f"🔍 RESPONSES API: Response received, type: {type(response)}")
-
-            # Extract fields from response
-            output_text = getattr(response, "output_text", getattr(response, "content", ""))
-            reasoning_obj = getattr(response, "reasoning", None)
-            response_id = getattr(response, "id", None)
-
-            # Debug reasoning type (only first time)
-            if reasoning_obj and not hasattr(self, "_reasoning_logged"):
-                print(f"🔍 RESPONSES API: Reasoning type: {type(reasoning_obj)}")
-                print(
-                    f"🔍 RESPONSES API: Reasoning attributes: {[x for x in dir(reasoning_obj) if not x.startswith('_')]}"
-                )
-                self._reasoning_logged = True
-
-            # Handle reasoning - it might be an object or a string
-            reasoning = None
-            if reasoning_obj:
-                if isinstance(reasoning_obj, str):
-                    # Synth backend returns full reasoning as string
-                    reasoning = reasoning_obj
-                else:
-                    # OpenAI returns a Reasoning object
-                    # Try to get summary first, but preserve entire object if no summary
-                    if hasattr(reasoning_obj, "summary") and reasoning_obj.summary:
-                        reasoning = reasoning_obj.summary
-                    else:
-                        # Preserve the full object structure as JSON
-                        # This includes effort level and any other fields
-                        if hasattr(reasoning_obj, "model_dump_json"):
-                            reasoning = reasoning_obj.model_dump_json()
-                        elif hasattr(reasoning_obj, "to_dict"):
-                            import json
-
-                            reasoning = json.dumps(reasoning_obj.to_dict())
-                        else:
-                            reasoning = str(reasoning_obj)
-
-            # Handle tool calls if present
-            tool_calls = None
-            if hasattr(response, "tool_calls") and response.tool_calls:
-                tool_calls = [
-                    {
-                        "id": tc.id,
-                        "type": tc.type,
-                        "function": {
-                            "name": tc.function.name,
-                            "arguments": tc.function.arguments,
-                        },
-                    }
-                    for tc in response.tool_calls
-                ]
-
-            print(f"🔍 RESPONSES API: Extracted response_id = {response_id}")
-
-            return BaseLMResponse(
-                raw_response=output_text,
-                response_id=response_id,
-                reasoning=reasoning,
-                api_type="responses",
-                tool_calls=tool_calls,
-            )
-
-        except (AttributeError, Exception) as e:
-            print(f"🔍 RESPONSES API: Error calling Responses API: {e}")
-            # No fallback - raise the error
-            raise
-
-    async def _hit_api_async_harmony(
-        self,
-        model: str,
-        messages: list[dict[str, Any]],
-        lm_config: dict[str, Any],
-        previous_response_id: str | None = None,
-        use_ephemeral_cache_only: bool = False,
-        tools: list[BaseTool] | None = None,
-    ) -> BaseLMResponse:
-        """Use Harmony encoding for OSS-GPT models."""
-        if not self.harmony_available:
-            raise ImportError(
-                "openai-harmony package required for OSS-GPT models. Install with: pip install openai-harmony"
-            )
-
-        from openai_harmony import Conversation, Message, Role
-
-        # Convert messages to Harmony format
-        harmony_messages = []
-        for msg in messages:
-            role = (
-                Role.SYSTEM
-                if msg["role"] == "system"
-                else (Role.USER if msg["role"] == "user" else Role.ASSISTANT)
-            )
-            content = msg["content"]
-            # Handle multimodal content
-            if isinstance(content, list):
-                # Extract text content for now
-                text_parts = [
-                    part.get("text", "") for part in content if part.get("type") == "text"
-                ]
-                content = " ".join(text_parts)
-            harmony_messages.append(Message.from_role_and_content(role, content))
-
-        conv = Conversation.from_messages(harmony_messages)
-        tokens = self.harmony_enc.render_conversation_for_completion(conv, Role.ASSISTANT)
-
-        # For now, we'll need to integrate with Synth GPU endpoint
-        # This would require the actual endpoint to be configured
-        # Placeholder for actual Synth GPU call
-        import os
-
-        import aiohttp
-
-        synth_gpu_endpoint = os.getenv("SYNTH_GPU_HARMONY_ENDPOINT")
-        if not synth_gpu_endpoint:
-            raise ValueError("SYNTH_GPU_HARMONY_ENDPOINT environment variable not set")
-
-        async with (
-            aiohttp.ClientSession() as session,
-            session.post(
-                f"{synth_gpu_endpoint}/v1/completions",
-                json={
-                    "model": model,
-                    "prompt": tokens,
-                    "max_tokens": lm_config.get("max_tokens", 4096),
-                    "temperature": lm_config.get("temperature", 0.8),
-                },
-            ) as resp,
-        ):
-            result = await resp.json()
-
-        # Parse response using Harmony
-        response_tokens = result.get("choices", [{}])[0].get("text", "")
-        parsed = self.harmony_enc.parse_messages_from_completion_tokens(
-            response_tokens, Role.ASSISTANT
-        )
-
-        if parsed:
-            assistant_msg = (
-                parsed[-1].content_text()
-                if hasattr(parsed[-1], "content_text")
-                else str(parsed[-1])
-            )
-        else:
-            assistant_msg = response_tokens
-
-        return BaseLMResponse(
-            raw_response=assistant_msg,
-            response_id=previous_response_id or str(uuid.uuid4()),
-            api_type="harmony",
-        )

synth_ai/v0/lm/vendors/retries.py

@@ -1,22 +0,0 @@
-import os
-
-import backoff
-
-# Number of retry attempts that some legacy decorators rely on.
-BACKOFF_TOLERANCE: int = 20
-
-# Maximum wall-clock seconds allowed for exponential back-off when retrying an
-# LLM API call.  This can be overridden at runtime with the environment
-# variable `SYNTH_AI_MAX_BACKOFF`.
-
-try:
-    MAX_BACKOFF: int = max(1, int(os.getenv("SYNTH_AI_MAX_BACKOFF", "120")))
-except ValueError:
-    MAX_BACKOFF = 120
-
-# Re-export backoff for convenient import patterns elsewhere
-__all__ = [
-    "BACKOFF_TOLERANCE",
-    "MAX_BACKOFF",
-    "backoff",
-]

synth_ai/v0/lm/vendors/supported/custom_endpoint.py

@@ -1,415 +0,0 @@
-import asyncio
-import json
-import os
-import random
-import re
-import time
-from typing import Any
-
-import httpx
-import requests
-from requests.adapters import HTTPAdapter
-from urllib3.util.retry import Retry
-
-from synth_ai.v0.lm.caching.initialize import get_cache_handler
-from synth_ai.v0.lm.tools.base import BaseTool
-from synth_ai.v0.lm.vendors.base import BaseLMResponse, VendorBase
-
-# Exception types for retry
-CUSTOM_ENDPOINT_EXCEPTIONS_TO_RETRY: tuple[type[Exception], ...] = (
-    requests.RequestException,
-    requests.Timeout,
-    httpx.RequestError,
-    httpx.TimeoutException,
-)
-
-
-class CustomEndpointAPI(VendorBase):
-    """Generic vendor client for custom OpenAI-compatible endpoints."""
-
-    used_for_structured_outputs: bool = False
-    exceptions_to_retry: list = list(CUSTOM_ENDPOINT_EXCEPTIONS_TO_RETRY)
-
-    def __init__(self, endpoint_url: str):
-        # Validate and sanitize URL
-        self._validate_endpoint_url(endpoint_url)
-        self.endpoint_url = endpoint_url
-
-        # Construct full chat completions URL
-        if endpoint_url.endswith("/"):
-            endpoint_url = endpoint_url[:-1]
-
-        # Handle full URLs that already include protocol
-        if endpoint_url.startswith(("http://", "https://")):
-            # Remove protocol and domain part, keep only the base path if any
-            parsed = endpoint_url.replace("https://", "").replace("http://", "")
-            base_url = parsed.split("/")[0]  # Get domain only
-            self.chat_completions_url = f"https://{base_url}/chat/completions"
-            self.health_url = f"https://{base_url}/health"
-        else:
-            # Original logic for domain-only URLs
-            self.chat_completions_url = f"https://{endpoint_url}/chat/completions"
-            self.health_url = f"https://{endpoint_url}/health"
-
-        # Setup session with connection pooling and retries
-        self.session = self._create_session()
-        self.async_client = None  # Lazy init
-
-        # Get auth token from environment (generic support for any auth)
-        self.auth_token = os.environ.get("CUSTOM_ENDPOINT_API_TOKEN")
-
-    def _validate_endpoint_url(self, url: str) -> None:
-        """Validate endpoint URL format and prevent SSRF."""
-        # Block dangerous URL patterns
-        dangerous_patterns = [
-            "file://",
-            "ftp://",
-            "gopher://",
-            "localhost",
-            "127.",
-            "0.0.0.0",
-            "10.",
-            "192.168.",
-            "172.16.",
-            "172.17.",
-            "172.18.",
-            "172.19.",
-            "172.20.",
-            "172.21.",
-            "172.22.",
-            "172.23.",
-            "172.24.",
-            "172.25.",
-            "172.26.",
-            "172.27.",
-            "172.28.",
-            "172.29.",
-            "172.30.",
-            "172.31.",
-            "169.254.",  # link-local
-            "::1",
-            "fc00:",
-            "fd00:",
-            "fe80:",  # IPv6 private
-        ]
-
-        for pattern in dangerous_patterns:
-            if pattern in url.lower():
-                raise ValueError(f"Blocked URL pattern for security: {pattern}")
-
-        # Limit URL length
-        if len(url) > 256:
-            raise ValueError("Endpoint URL too long (max 256 chars)")
-
-        # Basic URL format check
-        if not re.match(r"^[a-zA-Z0-9\-._~:/?#\[\]@!$&\'()*+,;=]+$", url):
-            raise ValueError(f"Invalid URL format: {url}")
-
-    def _create_session(self) -> requests.Session:
-        """Create session with retry strategy and connection pooling."""
-        session = requests.Session()
-
-        # Exponential backoff with jitter
-        retry_strategy = Retry(
-            total=3,
-            backoff_factor=1,
-            status_forcelist=[429, 500, 502, 503, 504],
-            allowed_methods=["POST", "GET"],
-        )
-
-        adapter = HTTPAdapter(max_retries=retry_strategy, pool_connections=10, pool_maxsize=20)
-
-        session.mount("http://", adapter)
-        session.mount("https://", adapter)
-
-        return session
-
-    async def _get_async_client(self) -> httpx.AsyncClient:
-        """Lazy init async client with shared retry logic."""
-        if self.async_client is None:
-            self.async_client = httpx.AsyncClient(
-                timeout=httpx.Timeout(30.0),
-                limits=httpx.Limits(max_connections=10, max_keepalive_connections=5),
-            )
-        return self.async_client
-
-    def _get_timeout(self, lm_config: dict[str, Any]) -> float:
-        """Get timeout with per-call override support."""
-        return lm_config.get(
-            "timeout", float(os.environ.get("CUSTOM_ENDPOINT_REQUEST_TIMEOUT", "30"))
-        )
-
-    def _get_temperature_override(self) -> float | None:
-        """Get temperature override from environment for this specific endpoint."""
-        # Create a safe env var key from the endpoint URL
-        # e.g., "example.com/api" -> "CUSTOM_ENDPOINT_TEMP_EXAMPLE_COM_API"
-        safe_key = re.sub(r"[^A-Za-z0-9]", "_", self.endpoint_url).upper()
-        safe_key = safe_key[:64]  # Limit length
-
-        env_key = f"CUSTOM_ENDPOINT_TEMP_{safe_key}"
-        temp_str = os.environ.get(env_key)
-        return float(temp_str) if temp_str else None
-
-    def _compress_tool_schema(self, schema: dict[str, Any]) -> dict[str, Any]:
-        """Compress JSON schema to reduce token usage."""
-        if isinstance(schema, dict):
-            # Remove verbose keys
-            compressed = {
-                k: self._compress_tool_schema(v)
-                for k, v in schema.items()
-                if k not in ["title", "$ref", "$schema"]
-            }
-            # Shorten descriptions
-            if "description" in compressed and len(compressed["description"]) > 50:
-                compressed["description"] = compressed["description"][:47] + "..."
-            return compressed
-        elif isinstance(schema, list):
-            return [self._compress_tool_schema(item) for item in schema]
-        return schema
-
-    def _inject_tools_into_prompt(self, system_message: str, tools: list[BaseTool]) -> str:
-        """Inject tool definitions with compressed schemas and clear output format."""
-        if not tools:
-            return system_message
-
-        tool_descriptions = []
-        for tool in tools:
-            schema = tool.arguments.model_json_schema()
-            compressed_schema = self._compress_tool_schema(schema)
-
-            tool_desc = f"Tool: {tool.name}\nDesc: {tool.description}\nParams: {json.dumps(compressed_schema, separators=(',', ':'))}"
-            tool_descriptions.append(tool_desc)
-
-        tools_text = "\n".join(tool_descriptions)
-
-        return f"""{system_message}
-
-Available tools:
-{tools_text}
-
-IMPORTANT: To use a tool, respond with JSON wrapped in ```json fences:
-```json
-{{"tool_call": {{"name": "tool_name", "arguments": {{...}}}}}}
-```
-
-For regular responses, just respond normally without JSON fences."""
-
-    def _extract_tool_calls(self, content: str, tools: list[BaseTool]) -> tuple[list | None, str]:
-        """Extract and validate tool calls from response."""
-        # Look for JSON fenced blocks
-        json_pattern = r"```json\s*(\{.*?\})\s*```"
-        matches = re.findall(json_pattern, content, re.DOTALL)
-
-        if not matches:
-            return None, content
-
-        tool_calls = []
-        cleaned_content = content
-
-        for match in matches:
-            try:
-                tool_data = json.loads(match)
-                if "tool_call" in tool_data:
-                    call_data = tool_data["tool_call"]
-                    tool_name = call_data.get("name")
-
-                    # Validate against available tools
-                    matching_tool = next((t for t in tools if t.name == tool_name), None)
-                    if matching_tool:
-                        # Validate arguments with pydantic
-                        validated_args = matching_tool.arguments(**call_data.get("arguments", {}))
-                        tool_calls.append(
-                            {"name": tool_name, "arguments": validated_args.model_dump()}
-                        )
-
-                        # Remove tool call from content
-                        cleaned_content = cleaned_content.replace(
-                            f"```json\n{match}\n```", ""
-                        ).strip()
-
-            except (json.JSONDecodeError, Exception):
-                # Fall back to treating as normal text if validation fails
-                continue
-
-        return tool_calls if tool_calls else None, cleaned_content
-
-    def _exponential_backoff_with_jitter(self, attempt: int) -> float:
-        """Calculate backoff time with jitter to prevent thundering herd."""
-        base_delay = min(2**attempt, 32)  # Cap at 32 seconds
-        jitter = random.uniform(0, 1)
-        return base_delay + jitter
-
-    def _handle_rate_limit(self, response: requests.Response) -> None:
-        """Extract and propagate rate limit information."""
-        if response.status_code == 429:
-            retry_after = response.headers.get("Retry-After")
-            if retry_after:
-                # Bubble up to synth-ai scheduler
-                raise requests.exceptions.RetryError(f"Rate limited. Retry after {retry_after}s")
-
-    async def _hit_api_async(
-        self,
-        model: str,
-        messages: list[dict[str, Any]],
-        lm_config: dict[str, Any],
-        use_ephemeral_cache_only: bool = False,
-        reasoning_effort: str = "low",
-        tools: list[BaseTool] | None = None,
-    ) -> BaseLMResponse:
-        """Async API call with comprehensive error handling and streaming support."""
-
-        # Cache integration - check first
-        used_cache_handler = get_cache_handler(use_ephemeral_cache_only)
-        cache_result = used_cache_handler.hit_managed_cache(
-            model, messages, lm_config=lm_config, tools=tools
-        )
-        if cache_result:
-            return cache_result
-
-        # Apply tool injection
-        if tools and messages:
-            messages = messages.copy()
-            if messages and messages[0].get("role") == "system":
-                messages[0]["content"] = self._inject_tools_into_prompt(
-                    messages[0]["content"], tools
-                )
-
-        # Prepare request
-        headers = {"Content-Type": "application/json"}
-        if self.auth_token:
-            headers["Authorization"] = f"Bearer {self.auth_token}"
-
-        # Apply temperature override
-        temp_override = self._get_temperature_override()
-        request_temp = temp_override if temp_override else lm_config.get("temperature", 0.7)
-
-        payload = {
-            "model": model,  # Pass through the model name
-            "messages": messages,
-            "temperature": request_temp,
-            "stream": lm_config.get("stream", False),
-        }
-
-        timeout = self._get_timeout(lm_config)
-        client = await self._get_async_client()
-
-        # Make request with retry logic
-        for attempt in range(3):
-            try:
-                response = await client.post(
-                    self.chat_completions_url, json=payload, headers=headers, timeout=timeout
-                )
-
-                if response.status_code == 429:
-                    self._handle_rate_limit(response)
-
-                response.raise_for_status()
-
-                response_data = response.json()
-                content = response_data["choices"][0]["message"]["content"]
-
-                # Extract tool calls
-                tool_calls, clean_content = self._extract_tool_calls(content, tools or [])
-
-                lm_response = BaseLMResponse(
-                    raw_response=clean_content, structured_output=None, tool_calls=tool_calls
-                )
-
-                # Add to cache
-                used_cache_handler.add_to_managed_cache(
-                    model, messages, lm_config=lm_config, output=lm_response, tools=tools
-                )
-
-                return lm_response
-
-            except (httpx.RequestError, httpx.TimeoutException):
-                if attempt == 2:  # Last attempt
-                    raise
-                await asyncio.sleep(self._exponential_backoff_with_jitter(attempt))
-
-    def _hit_api_sync(
-        self,
-        model: str,
-        messages: list[dict[str, Any]],
-        lm_config: dict[str, Any],
-        use_ephemeral_cache_only: bool = False,
-        reasoning_effort: str = "low",
-        tools: list[BaseTool] | None = None,
-    ) -> BaseLMResponse:
-        """Sync version with same logic as async."""
-
-        # Cache integration - check first
-        used_cache_handler = get_cache_handler(use_ephemeral_cache_only)
-        cache_result = used_cache_handler.hit_managed_cache(
-            model, messages, lm_config=lm_config, tools=tools
-        )
-        if cache_result:
-            return cache_result
-
-        # Apply tool injection
-        if tools and messages:
-            messages = messages.copy()
-            if messages and messages[0].get("role") == "system":
-                messages[0]["content"] = self._inject_tools_into_prompt(
-                    messages[0]["content"], tools
-                )
-
-        # Prepare request
-        headers = {"Content-Type": "application/json"}
-        if self.auth_token:
-            headers["Authorization"] = f"Bearer {self.auth_token}"
-
-        # Apply temperature override
-        temp_override = self._get_temperature_override()
-        request_temp = temp_override if temp_override else lm_config.get("temperature", 0.7)
-
-        payload = {
-            "model": model,  # Pass through the model name
-            "messages": messages,
-            "temperature": request_temp,
-            "stream": lm_config.get("stream", False),
-        }
-
-        timeout = self._get_timeout(lm_config)
-
-        # Make request with retry logic
-        for attempt in range(3):
-            try:
-                response = self.session.post(
-                    self.chat_completions_url, json=payload, headers=headers, timeout=timeout
-                )
-
-                if response.status_code == 429:
-                    self._handle_rate_limit(response)
-
-                response.raise_for_status()
-
-                response_data = response.json()
-                content = response_data["choices"][0]["message"]["content"]
-
-                # Extract tool calls
-                tool_calls, clean_content = self._extract_tool_calls(content, tools or [])
-
-                lm_response = BaseLMResponse(
-                    raw_response=clean_content, structured_output=None, tool_calls=tool_calls
-                )
-
-                # Add to cache
-                used_cache_handler.add_to_managed_cache(
-                    model, messages, lm_config=lm_config, output=lm_response, tools=tools
-                )
-
-                return lm_response
-
-            except (requests.RequestException, requests.Timeout):
-                if attempt == 2:  # Last attempt
-                    raise
-                time.sleep(self._exponential_backoff_with_jitter(attempt))
-
-    def __del__(self):
-        """Cleanup resources."""
-        if hasattr(self, "session"):
-            self.session.close()
-        if hasattr(self, "async_client") and self.async_client:
-            # Schedule cleanup for async client
-            pass