flowent 0.0.1 → 0.0.5

package/backend/src/flowent/providers/openai_responses.py

@@ -0,0 +1,655 @@
+from __future__ import annotations
+
+import base64
+import json
+import time
+from collections.abc import Callable
+from typing import Any
+
+from loguru import logger
+
+from flowent.model_metadata import build_model_info
+from flowent.models import (
+    LLMOutputImagePart,
+    LLMOutputTextPart,
+    LLMResponse,
+    LLMUsage,
+    ModelInfo,
+)
+from flowent.models import ToolCallResult as ToolCall
+from flowent.network import (
+    RequestException,
+    create_http_session,
+    iter_response_lines,
+    read_response_text,
+    response_looks_like_html,
+    truncate_text,
+)
+from flowent.providers import LLMProvider
+from flowent.providers.content import (
+    collapse_parts_to_text,
+    to_openai_responses_content,
+)
+from flowent.providers.errors import (
+    build_access_blocked_error,
+    build_network_error,
+    build_status_error,
+)
+from flowent.providers.headers import merge_headers
+from flowent.settings import ModelParams
+
+REASONING_MODEL_PREFIXES = ("gpt-5", "o1", "o3", "o4")
+VERBOSITY_MODEL_PREFIXES = ("gpt-5",)
+REASONING_DELTA_EVENT_TYPES = {
+    "response.reasoning.delta",
+    "response.reasoning_text.delta",
+    "response.reasoning_summary_text.delta",
+}
+
+
+def _supports_reasoning(model: str) -> bool:
+    return model.startswith(REASONING_MODEL_PREFIXES)
+
+
+def _build_reasoning_config(
+    model: str,
+    model_params: ModelParams | None,
+) -> dict[str, str] | None:
+    if not _supports_reasoning(model) or model_params is None:
+        return None
+    effort = model_params.reasoning_effort
+    if effort is None or effort == "none":
+        return None
+
+    return {
+        "effort": effort,
+        "summary": "detailed",
+    }
+
+
+def _build_text_config(
+    model: str,
+    model_params: ModelParams | None,
+) -> dict[str, object] | None:
+    if model_params is None or not model.startswith(VERBOSITY_MODEL_PREFIXES):
+        return None
+    if not model_params.verbosity:
+        return None
+
+    return {
+        "format": {"type": "text"},
+        "verbosity": model_params.verbosity,
+    }
+
+
+def _extract_reasoning_text_from_item(item: dict[str, Any]) -> str | None:
+    parts: list[str] = []
+
+    for key in ("summary", "content"):
+        raw_value = item.get(key)
+
+        if isinstance(raw_value, str):
+            if raw_value.strip():
+                parts.append(raw_value)
+            continue
+
+        if not isinstance(raw_value, list):
+            continue
+
+        for entry in raw_value:
+            if not isinstance(entry, dict):
+                continue
+            text = entry.get("text")
+            if isinstance(text, str) and text.strip():
+                parts.append(text)
+
+    if not parts:
+        return None
+
+    return "".join(parts)
+
+
+def _extract_reasoning_text_from_output(output: Any) -> str | None:
+    if not isinstance(output, list):
+        return None
+
+    parts: list[str] = []
+    for item in output:
+        if not isinstance(item, dict):
+            continue
+        if item.get("type") != "reasoning":
+            continue
+        text = _extract_reasoning_text_from_item(item)
+        if text:
+            parts.append(text)
+
+    if not parts:
+        return None
+
+    return "\n\n".join(parts)
+
+
+def _extract_reasoning_tokens(response: dict[str, Any]) -> int:
+    usage = response.get("usage")
+    if not isinstance(usage, dict):
+        return 0
+
+    output_tokens_details = usage.get("output_tokens_details")
+    if not isinstance(output_tokens_details, dict):
+        return 0
+
+    tokens = output_tokens_details.get("reasoning_tokens")
+    return tokens if isinstance(tokens, int) and tokens > 0 else 0
+
+
+def _collect_usage_details(
+    value: dict[str, Any],
+    *,
+    prefix: str = "",
+) -> dict[str, int]:
+    details: dict[str, int] = {}
+    for key, raw_item in value.items():
+        detail_key = f"{prefix}.{key}" if prefix else key
+        if isinstance(raw_item, bool):
+            continue
+        if isinstance(raw_item, int) and raw_item >= 0:
+            details[detail_key] = raw_item
+            continue
+        if isinstance(raw_item, dict):
+            details.update(_collect_usage_details(raw_item, prefix=detail_key))
+    return details
+
+
+def _extract_usage(response: dict[str, Any]) -> LLMUsage | None:
+    usage = response.get("usage")
+    if not isinstance(usage, dict):
+        return None
+
+    total_tokens = usage.get("total_tokens")
+    if isinstance(total_tokens, bool) or not isinstance(total_tokens, int):
+        return None
+
+    input_tokens = usage.get("input_tokens")
+    if isinstance(input_tokens, bool) or not isinstance(input_tokens, int):
+        input_tokens = None
+
+    output_tokens = usage.get("output_tokens")
+    if isinstance(output_tokens, bool) or not isinstance(output_tokens, int):
+        output_tokens = None
+
+    cached_input_tokens = usage.get("cached_input_tokens")
+    if isinstance(cached_input_tokens, bool) or not isinstance(
+        cached_input_tokens, int
+    ):
+        cached_input_tokens = None
+    input_token_details = usage.get("input_tokens_details")
+    if (
+        cached_input_tokens is None
+        and isinstance(input_token_details, dict)
+        and isinstance(input_token_details.get("cached_tokens"), int)
+    ):
+        cached_input_tokens = input_token_details["cached_tokens"]
+
+    details = _collect_usage_details(usage)
+    for key in (
+        "total_tokens",
+        "input_tokens",
+        "output_tokens",
+        "cached_input_tokens",
+        "input_tokens_details.cached_tokens",
+    ):
+        details.pop(key, None)
+
+    return LLMUsage(
+        total_tokens=total_tokens,
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cached_input_tokens=cached_input_tokens,
+        cache_read_tokens=cached_input_tokens,
+        details=details,
+    )
+
+
+def _decode_output_image(item: dict[str, Any]) -> LLMOutputImagePart | None:
+    encoded_data = (
+        item.get("result") or item.get("image_base64") or item.get("b64_json")
+    )
+    if not isinstance(encoded_data, str) or not encoded_data:
+        return None
+    try:
+        image_data = base64.b64decode(encoded_data)
+    except ValueError:
+        return None
+    mime_type = item.get("mime_type")
+    return LLMOutputImagePart(
+        data=image_data,
+        mime_type=mime_type if isinstance(mime_type, str) else "image/png",
+    )
+
+
+def _extract_output_parts(output: Any) -> list[LLMOutputTextPart | LLMOutputImagePart]:
+    if not isinstance(output, list):
+        return []
+
+    parts: list[LLMOutputTextPart | LLMOutputImagePart] = []
+    for item in output:
+        if not isinstance(item, dict):
+            continue
+        item_type = item.get("type")
+        if item_type == "message":
+            content = item.get("content")
+            if not isinstance(content, list):
+                continue
+            for entry in content:
+                if not isinstance(entry, dict):
+                    continue
+                entry_type = entry.get("type")
+                text = entry.get("text")
+                if entry_type in {"output_text", "text"} and isinstance(text, str):
+                    parts.append(LLMOutputTextPart(text=text))
+        elif item_type == "image_generation_call":
+            image_part = _decode_output_image(item)
+            if image_part is not None:
+                parts.append(image_part)
+    return parts
+
+
+def _format_reasoning_fallback(reasoning_tokens: int) -> str:
+    if reasoning_tokens > 0:
+        return f"Internal reasoning · {reasoning_tokens}"
+
+    return "Internal reasoning"
+
+
+class OpenAIResponsesProvider(LLMProvider):
+    def __init__(
+        self,
+        provider_name: str,
+        api_base_url: str,
+        api_key: str = "",
+        headers: dict[str, str] | None = None,
+        model: str = "",
+        request_timeout_seconds: float = 120.0,
+    ) -> None:
+        self._provider_name = provider_name
+        self._api_base_url = api_base_url.rstrip("/")
+        self._api_key = api_key
+        self._header_overrides = dict(headers or {})
+        self._model = model
+        self._request_timeout_seconds = request_timeout_seconds
+        self._client = create_http_session(timeout=self._request_timeout_seconds)
+
+    def _headers(self) -> dict[str, str]:
+        headers: dict[str, str] = {"Content-Type": "application/json"}
+        if self._api_key:
+            headers["Authorization"] = f"Bearer {self._api_key}"
+        return merge_headers(headers, self._header_overrides)
+
+    def _convert_tools(self, tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        result = []
+        for t in tools:
+            fn = t.get("function", {})
+            result.append(
+                {
+                    "type": "function",
+                    "name": fn.get("name", ""),
+                    "description": fn.get("description", ""),
+                    "parameters": fn.get("parameters", {}),
+                }
+            )
+        return result
+
+    def _convert_messages(
+        self, messages: list[dict[str, Any]]
+    ) -> tuple[str | None, list[dict[str, Any]]]:
+        system_prompt: str | None = None
+        input_items: list[dict[str, Any]] = []
+
+        for msg in messages:
+            role = msg.get("role")
+            content = msg.get("content")
+
+            if role == "system":
+                system_prompt = content
+                continue
+
+            if role == "user":
+                input_items.append(
+                    {
+                        "role": "user",
+                        "content": to_openai_responses_content(
+                            content,
+                            allow_images=True,
+                        ),
+                    }
+                )
+            elif role == "assistant":
+                tool_calls = msg.get("tool_calls")
+                if tool_calls:
+                    for tc in tool_calls:
+                        fn = tc.get("function", {})
+                        input_items.append(
+                            {
+                                "type": "function_call",
+                                "call_id": tc.get("id", ""),
+                                "name": fn.get("name", ""),
+                                "arguments": fn.get("arguments", "{}"),
+                            }
+                        )
+                else:
+                    input_items.append(
+                        {
+                            "role": "assistant",
+                            "content": collapse_parts_to_text(content),
+                        }
+                    )
+            elif role == "tool":
+                input_items.append(
+                    {
+                        "type": "function_call_output",
+                        "call_id": msg.get("tool_call_id", ""),
+                        "output": content or "",
+                    }
+                )
+
+        return system_prompt, input_items
+
+    def _supports_output_image(self) -> bool:
+        model_info = build_model_info(
+            provider_type="openai_responses",
+            model_id=self._model,
+        )
+        return model_info.capabilities.output_image
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        on_chunk: Callable[[str, str], None] | None = None,
+        register_interrupt: Callable[[Callable[[], None] | None], None] | None = None,
+        model_params: ModelParams | None = None,
+    ) -> LLMResponse:
+        url = f"{self._api_base_url}/responses"
+        system_prompt, input_items = self._convert_messages(messages)
+
+        payload: dict[str, Any] = {
+            "model": self._model,
+            "input": input_items,
+            "stream": True,
+        }
+        reasoning = _build_reasoning_config(self._model, model_params)
+        if reasoning:
+            payload["reasoning"] = reasoning
+        text_config = _build_text_config(self._model, model_params)
+        if text_config:
+            payload["text"] = text_config
+        if model_params is not None:
+            if model_params.max_output_tokens is not None:
+                payload["max_output_tokens"] = model_params.max_output_tokens
+            if reasoning is None:
+                if model_params.temperature is not None:
+                    payload["temperature"] = model_params.temperature
+                if model_params.top_p is not None:
+                    payload["top_p"] = model_params.top_p
+        if system_prompt:
+            payload["instructions"] = system_prompt
+        payload_tools = []
+        if tools:
+            payload_tools.extend(self._convert_tools(tools))
+        if self._supports_output_image():
+            payload_tools.append({"type": "image_generation"})
+        if payload_tools:
+            payload["tools"] = payload_tools
+            payload["tool_choice"] = "auto"
+
+        logger.debug(
+            "[{}] OpenAI Responses chat request: model={}, input_items={}, tools={}",
+            self._provider_name,
+            self._model,
+            len(input_items),
+            len(tools) if tools else 0,
+        )
+
+        t0 = time.perf_counter()
+        content_parts: list[str] = []
+        thinking_parts: list[str] = []
+        response_parts: list[LLMOutputTextPart | LLMOutputImagePart] | None = None
+        tool_calls: list[ToolCall] = []
+        chunk_count = 0
+        saw_reasoning_item = False
+        saw_reasoning_text = False
+        reasoning_tokens = 0
+        response_usage: LLMUsage | None = None
+        raw_usage: dict[str, Any] | None = None
+
+        current_tool: dict[str, Any] = {}
+        client = self._client
+        if register_interrupt is not None:
+            register_interrupt(client.close)
+
+        try:
+            with client.stream(
+                "POST",
+                url,
+                headers=self._headers(),
+                json=payload,
+            ) as response:
+                if register_interrupt is not None:
+                    register_interrupt(response.close)
+                if response_looks_like_html(response):
+                    raise build_access_blocked_error(
+                        provider_name=self._provider_name,
+                        provider_type="openai_responses",
+                        model=self._model,
+                        base_url=self._api_base_url,
+                        status_code=response.status_code,
+                        detail=truncate_text(read_response_text(response)),
+                    )
+                if response.status_code != 200:
+                    body = truncate_text(read_response_text(response))
+                    elapsed = time.perf_counter() - t0
+                    logger.error(
+                        "LLM API error [provider={}, model={}, type=openai_responses]: {} - {} ({:.2f}s)",
+                        self._provider_name,
+                        self._model,
+                        response.status_code,
+                        body[:500],
+                        elapsed,
+                    )
+                    raise build_status_error(
+                        provider_name=self._provider_name,
+                        provider_type="openai_responses",
+                        model=self._model,
+                        base_url=self._api_base_url,
+                        status_code=response.status_code,
+                        body=body,
+                    )
+
+                for line in iter_response_lines(response):
+                    if not line or line.startswith(":"):
+                        continue
+                    if not line.startswith("data: "):
+                        continue
+                    data_str = line[6:]
+                    if data_str.strip() == "[DONE]":
+                        break
+
+                    try:
+                        event = json.loads(data_str)
+                    except json.JSONDecodeError:
+                        continue
+
+                    chunk_count += 1
+                    event_type = event.get("type", "")
+
+                    if event_type in REASONING_DELTA_EVENT_TYPES:
+                        text = event.get("delta", "")
+                        if text:
+                            saw_reasoning_text = True
+                            thinking_parts.append(text)
+                            if on_chunk:
+                                on_chunk("thinking", text)
+
+                    elif event_type == "response.output_text.delta":
+                        text = event.get("delta", "")
+                        if text:
+                            content_parts.append(text)
+                            if on_chunk:
+                                on_chunk("content", text)
+
+                    elif event_type == "response.function_call_arguments.delta":
+                        delta = event.get("delta", "")
+                        current_tool.setdefault("arguments", "")
+                        current_tool["arguments"] += delta
+
+                    elif event_type == "response.output_item.added":
+                        item = event.get("item", {})
+                        if item.get("type") == "reasoning":
+                            saw_reasoning_item = True
+                        elif item.get("type") == "function_call":
+                            current_tool = {
+                                "id": item.get("call_id", ""),
+                                "name": item.get("name", ""),
+                                "arguments": "",
+                            }
+
+                    elif event_type == "response.output_item.done":
+                        item = event.get("item", {})
+                        if item.get("type") == "reasoning":
+                            saw_reasoning_item = True
+                            if not saw_reasoning_text:
+                                reasoning_text = _extract_reasoning_text_from_item(item)
+                                if reasoning_text:
+                                    saw_reasoning_text = True
+                                    thinking_parts.append(reasoning_text)
+                                    if on_chunk:
+                                        on_chunk("thinking", reasoning_text)
+                        elif item.get("type") == "function_call":
+                            call_id = item.get("call_id") or current_tool.get("id", "")
+                            name = item.get("name") or current_tool.get("name", "")
+                            args_str = item.get("arguments") or current_tool.get(
+                                "arguments", "{}"
+                            )
+                            try:
+                                args = json.loads(args_str)
+                            except json.JSONDecodeError:
+                                args = {}
+                            tool_calls.append(
+                                ToolCall(id=call_id, name=name, arguments=args)
+                            )
+                            current_tool = {}
+
+                    elif event_type == "response.completed":
+                        response_data = event.get("response", {})
+                        if isinstance(response_data, dict):
+                            reasoning_tokens = _extract_reasoning_tokens(response_data)
+                            response_usage = _extract_usage(response_data)
+                            usage_payload = response_data.get("usage")
+                            if isinstance(usage_payload, dict):
+                                raw_usage = dict(usage_payload)
+                            if not saw_reasoning_text:
+                                reasoning_text = _extract_reasoning_text_from_output(
+                                    response_data.get("output")
+                                )
+                                if reasoning_text:
+                                    saw_reasoning_text = True
+                                    thinking_parts.append(reasoning_text)
+                                    if on_chunk:
+                                        on_chunk("thinking", reasoning_text)
+                            output_parts = _extract_output_parts(
+                                response_data.get("output")
+                            )
+                            if output_parts:
+                                response_parts = output_parts
+        except RequestException as exc:
+            elapsed = time.perf_counter() - t0
+            logger.warning(
+                "LLM API transport error [provider={}, model={}, type=openai_responses]: {} ({:.2f}s)",
+                self._provider_name,
+                self._model,
+                exc,
+                elapsed,
+            )
+            raise build_network_error(
+                provider_name=self._provider_name,
+                provider_type="openai_responses",
+                model=self._model,
+                base_url=self._api_base_url,
+                error=exc,
+            ) from exc
+        finally:
+            close_client = getattr(client, "close", None)
+            if callable(close_client):
+                close_client()
+                self._client = create_http_session(
+                    timeout=self._request_timeout_seconds
+                )
+
+        elapsed = time.perf_counter() - t0
+        content = "".join(content_parts) or None
+        thinking = "".join(thinking_parts) or None
+        if thinking is None and (saw_reasoning_item or reasoning_tokens > 0):
+            thinking = _format_reasoning_fallback(reasoning_tokens)
+
+        logger.debug(
+            "[{}] OpenAI Responses chat done: {:.2f}s, chunks={}, content_len={}, thinking_len={}, tool_calls={}",
+            self._provider_name,
+            elapsed,
+            chunk_count,
+            len(content) if content else 0,
+            len(thinking) if thinking else 0,
+            len(tool_calls),
+        )
+
+        if tool_calls:
+            return LLMResponse(
+                content=content,
+                parts=response_parts,
+                tool_calls=tool_calls,
+                thinking=thinking,
+                usage=response_usage,
+                raw_usage=raw_usage,
+            )
+
+        return LLMResponse(
+            content=content or "",
+            parts=response_parts,
+            thinking=thinking,
+            usage=response_usage,
+            raw_usage=raw_usage,
+        )
+
+    def list_models(
+        self,
+        register_interrupt: Callable[[Callable[[], None] | None], None] | None = None,
+    ) -> list[ModelInfo]:
+        url = f"{self._api_base_url}/models"
+        client = self._client
+        close_client = getattr(client, "close", None)
+        if register_interrupt is not None and callable(close_client):
+            register_interrupt(close_client)
+        try:
+            resp = client.get(url, headers=self._headers())
+            resp.raise_for_status()
+            data = resp.json()
+            models = data.get("data", [])
+            return [
+                build_model_info(
+                    provider_type="openai_responses",
+                    model_id=m["id"],
+                )
+                for m in models
+            ]
+        except Exception as e:
+            logger.error(
+                "Failed to list models [provider={}, type=openai_responses]: {}",
+                self._provider_name,
+                e,
+            )
+            return []
+        finally:
+            if callable(close_client):
+                close_client()
+                self._client = create_http_session(
+                    timeout=self._request_timeout_seconds
+                )