Files
oai-web/server/providers/openai_provider.py
2026-04-08 12:43:24 +02:00

232 lines
8.3 KiB
Python

"""
providers/openai_provider.py — Direct OpenAI provider.
Uses the official openai SDK pointing at api.openai.com (default base URL).
Tool schema conversion reuses the same Anthropic→OpenAI format translation
as the OpenRouter provider (they share the same wire format).
"""
from __future__ import annotations
import json
import logging
from typing import Any
from openai import OpenAI, AsyncOpenAI
from .base import AIProvider, ProviderResponse, ToolCallResult, UsageStats
logger = logging.getLogger(__name__)
DEFAULT_MODEL = "gpt-4o"
# Models that use max_completion_tokens instead of max_tokens, and don't support
# tool_choice="auto" (reasoning models use implicit tool choice).
_REASONING_MODELS = frozenset({"o1", "o1-mini", "o1-preview"})
def _convert_content_blocks(blocks: list[dict]) -> list[dict]:
"""Convert Anthropic-native content blocks to OpenAI image_url format."""
result = []
for block in blocks:
if block.get("type") == "image":
src = block.get("source", {})
if src.get("type") == "base64":
data_url = f"data:{src['media_type']};base64,{src['data']}"
result.append({"type": "image_url", "image_url": {"url": data_url}})
else:
result.append(block)
return result
class OpenAIProvider(AIProvider):
def __init__(self, api_key: str) -> None:
self._client = OpenAI(api_key=api_key)
self._async_client = AsyncOpenAI(api_key=api_key)
@property
def name(self) -> str:
return "OpenAI"
@property
def default_model(self) -> str:
return DEFAULT_MODEL
# ── Public interface ──────────────────────────────────────────────────────
def chat(
self,
messages: list[dict],
tools: list[dict] | None = None,
system: str = "",
model: str = "",
max_tokens: int = 4096,
) -> ProviderResponse:
params = self._build_params(messages, tools, system, model, max_tokens)
try:
response = self._client.chat.completions.create(**params)
return self._parse_response(response)
except Exception as e:
logger.error(f"OpenAI chat error: {e}")
return ProviderResponse(text=f"Error: {e}", finish_reason="error")
async def chat_async(
self,
messages: list[dict],
tools: list[dict] | None = None,
system: str = "",
model: str = "",
max_tokens: int = 4096,
) -> ProviderResponse:
params = self._build_params(messages, tools, system, model, max_tokens)
try:
response = await self._async_client.chat.completions.create(**params)
return self._parse_response(response)
except Exception as e:
logger.error(f"OpenAI async chat error: {e}")
return ProviderResponse(text=f"Error: {e}", finish_reason="error")
# ── Internal helpers ──────────────────────────────────────────────────────
def _build_params(
self,
messages: list[dict],
tools: list[dict] | None,
system: str,
model: str,
max_tokens: int,
) -> dict:
model = model or self.default_model
openai_messages = self._convert_messages(messages, system, model)
params: dict = {
"model": model,
"messages": openai_messages,
}
is_reasoning = model in _REASONING_MODELS
if is_reasoning:
params["max_completion_tokens"] = max_tokens
else:
params["max_tokens"] = max_tokens
if tools:
params["tools"] = [self._to_openai_tool(t) for t in tools]
if not is_reasoning:
params["tool_choice"] = "auto"
return params
def _convert_messages(self, messages: list[dict], system: str, model: str) -> list[dict]:
"""Convert aide's internal message list to OpenAI format."""
result: list[dict] = []
# Reasoning models (o1, o1-mini) don't support system role — use user role instead
is_reasoning = model in _REASONING_MODELS
if system:
if is_reasoning:
result.append({"role": "user", "content": f"[System instructions]\n{system}"})
else:
result.append({"role": "system", "content": system})
i = 0
while i < len(messages):
msg = messages[i]
role = msg["role"]
if role == "system":
i += 1
continue # Already prepended above
if role == "assistant" and msg.get("tool_calls"):
openai_tool_calls = []
for tc in msg["tool_calls"]:
openai_tool_calls.append({
"id": tc["id"],
"type": "function",
"function": {
"name": tc["name"],
"arguments": json.dumps(tc["arguments"]),
},
})
out: dict[str, Any] = {"role": "assistant", "tool_calls": openai_tool_calls}
if msg.get("content"):
out["content"] = msg["content"]
result.append(out)
elif role == "tool":
# Group consecutive tool results; collect image blocks for injection
pending_images: list[dict] = []
while i < len(messages) and messages[i]["role"] == "tool":
t = messages[i]
content = t.get("content", "")
if isinstance(content, list):
text = " ".join(b.get("text", "") for b in content if b.get("type") == "text") or "[image]"
pending_images.extend(b for b in content if b.get("type") == "image")
content = text
result.append({"role": "tool", "tool_call_id": t["tool_call_id"], "content": content})
i += 1
if pending_images:
result.append({"role": "user", "content": _convert_content_blocks(pending_images)})
continue # i already advanced
else:
content = msg.get("content", "")
if isinstance(content, list):
content = _convert_content_blocks(content)
result.append({"role": role, "content": content})
i += 1
return result
@staticmethod
def _to_openai_tool(aide_tool: dict) -> dict:
"""Convert aide's Anthropic-native tool schema to OpenAI function-calling format."""
return {
"type": "function",
"function": {
"name": aide_tool["name"],
"description": aide_tool.get("description", ""),
"parameters": aide_tool.get("input_schema", {"type": "object", "properties": {}}),
},
}
def _parse_response(self, response) -> ProviderResponse:
choice = response.choices[0] if response.choices else None
if not choice:
return ProviderResponse(text=None, finish_reason="error")
message = choice.message
text = message.content or None
tool_calls: list[ToolCallResult] = []
if message.tool_calls:
for tc in message.tool_calls:
try:
arguments = json.loads(tc.function.arguments)
except json.JSONDecodeError:
arguments = {"_raw": tc.function.arguments}
tool_calls.append(ToolCallResult(
id=tc.id,
name=tc.function.name,
arguments=arguments,
))
usage = UsageStats()
if response.usage:
usage = UsageStats(
input_tokens=response.usage.prompt_tokens,
output_tokens=response.usage.completion_tokens,
)
finish_reason = choice.finish_reason or "stop"
if tool_calls:
finish_reason = "tool_use"
return ProviderResponse(
text=text,
tool_calls=tool_calls,
usage=usage,
finish_reason=finish_reason,
model=response.model,
)