first commit
This commit is contained in:
Hermes Agent
5
sn-image-base/scripts/sn_image_base/llm/__init__.py
Normal file
5
sn-image-base/scripts/sn_image_base/llm/__init__.py
Normal file
@@ -0,0 +1,5 @@
|
||||
# llm module - Language Model (text only)
|
||||
from .anthropic_adapter import AnthropicMessagesAdapter
|
||||
from .chat_completions_adapter import OpenAIChatAdapter
|
||||
|
||||
__all__ = ["AnthropicMessagesAdapter", "OpenAIChatAdapter"]
|
||||
161
sn-image-base/scripts/sn_image_base/llm/anthropic_adapter.py
Normal file
161
sn-image-base/scripts/sn_image_base/llm/anthropic_adapter.py
Normal file
@@ -0,0 +1,161 @@
|
||||
"""Anthropic Messages API adapter for text and vision."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from typing import Any
|
||||
|
||||
import httpx
|
||||
|
||||
from sn_image_base.utils.error_utils import U1HttpResponseParseError
|
||||
from sn_image_base.utils.httpx_client import httpx_response_raise_for_status_code
|
||||
from sn_image_base.vlm.utils import image_to_base64
|
||||
from sn_image_base.vlm.vlm_adapter import VlmAdapter
|
||||
|
||||
from .llm_adapter import LlmAdapter
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
DEFAULT_REQUEST_TIMEOUT = 150.0
|
||||
DEFAULT_MAX_TOKENS = 4096
|
||||
|
||||
|
||||
class AnthropicMessagesAdapter(LlmAdapter, VlmAdapter):
|
||||
"""Anthropic Messages API adapter for text-only and vision calls."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
endpoint_url: str,
|
||||
api_key: str,
|
||||
model: str,
|
||||
*,
|
||||
max_tokens: int = DEFAULT_MAX_TOKENS,
|
||||
timeout: float = DEFAULT_REQUEST_TIMEOUT,
|
||||
async_client: httpx.AsyncClient | None = None,
|
||||
) -> None:
|
||||
self._url = endpoint_url
|
||||
self._api_key = api_key
|
||||
self._default_model = model
|
||||
self._max_tokens = max_tokens
|
||||
self._timeout = timeout
|
||||
self._external_client = async_client
|
||||
self._client: httpx.AsyncClient | None = async_client
|
||||
logger.info(
|
||||
"AnthropicMessagesAdapter: endpoint=%s model=%s max_tokens=%s",
|
||||
self._url,
|
||||
self._default_model,
|
||||
self._max_tokens,
|
||||
)
|
||||
|
||||
def _get_client(self) -> httpx.AsyncClient:
|
||||
if self._client is None:
|
||||
self._client = httpx.AsyncClient(timeout=self._timeout)
|
||||
return self._client
|
||||
|
||||
@property
|
||||
def _headers(self) -> dict[str, str]:
|
||||
return {
|
||||
"Authorization": f"Bearer {self._api_key}",
|
||||
"Content-Type": "application/json",
|
||||
"x-api-key": self._api_key,
|
||||
}
|
||||
|
||||
@staticmethod
|
||||
def _build_vision_content(
|
||||
user_prompt: str,
|
||||
images: list[str | bytes],
|
||||
) -> list[dict[str, Any]]:
|
||||
blocks: list[dict[str, Any]] = [{"type": "text", "text": user_prompt}]
|
||||
for image in images:
|
||||
mime, b64 = image_to_base64(image)
|
||||
blocks.append(
|
||||
{
|
||||
"type": "image",
|
||||
"source": {
|
||||
"type": "base64",
|
||||
"media_type": mime,
|
||||
"data": b64,
|
||||
},
|
||||
}
|
||||
)
|
||||
return blocks
|
||||
|
||||
def _build_payload(
|
||||
self,
|
||||
user_prompt: str,
|
||||
system_prompt: str,
|
||||
model: str | None,
|
||||
*,
|
||||
images: list[str | bytes] | None = None,
|
||||
) -> dict[str, Any]:
|
||||
messages: list[dict[str, Any]] = []
|
||||
if system_prompt:
|
||||
messages.append({"role": "user", "content": system_prompt})
|
||||
|
||||
user_content: str | list[dict[str, Any]]
|
||||
if images:
|
||||
user_content = self._build_vision_content(user_prompt, images)
|
||||
else:
|
||||
user_content = user_prompt
|
||||
messages.append({"role": "user", "content": user_content})
|
||||
|
||||
return {
|
||||
"model": model or self._default_model,
|
||||
"messages": messages,
|
||||
"max_tokens": self._max_tokens,
|
||||
}
|
||||
|
||||
@staticmethod
|
||||
def _parse_response(data: dict[str, Any]) -> str:
|
||||
content = data.get("content", [])
|
||||
if content:
|
||||
for block in content:
|
||||
if isinstance(block, dict) and block.get("type") == "text":
|
||||
return block.get("text", "")
|
||||
|
||||
thinking = data.get("thinking")
|
||||
if thinking:
|
||||
return f"[Think] {thinking}"
|
||||
|
||||
raise RuntimeError("Anthropic Messages response has no extractable content.")
|
||||
|
||||
async def _post_payload(self, payload: dict[str, Any]) -> str:
|
||||
resp = await self._get_client().post(self._url, json=payload, headers=self._headers)
|
||||
httpx_response_raise_for_status_code(resp)
|
||||
try:
|
||||
data = resp.json()
|
||||
except ValueError as exc:
|
||||
raise U1HttpResponseParseError(
|
||||
detail=f"Failed to parse HTTP response. {resp.request.url}. Response content: {resp.content}",
|
||||
code=resp.status_code,
|
||||
) from exc
|
||||
return self._parse_response(data)
|
||||
|
||||
async def text_completion(
|
||||
self,
|
||||
user_prompt: str,
|
||||
system_prompt: str = "",
|
||||
model: str | None = None,
|
||||
) -> str:
|
||||
payload = self._build_payload(user_prompt, system_prompt, model)
|
||||
return await self._post_payload(payload)
|
||||
|
||||
async def vision_completion(
|
||||
self,
|
||||
user_prompt: str,
|
||||
images: list[str | bytes],
|
||||
system_prompt: str = "",
|
||||
model: str | None = None,
|
||||
) -> str:
|
||||
payload = self._build_payload(
|
||||
user_prompt,
|
||||
system_prompt,
|
||||
model,
|
||||
images=images,
|
||||
)
|
||||
return await self._post_payload(payload)
|
||||
|
||||
async def aclose(self) -> None:
|
||||
if self._external_client is None and self._client is not None:
|
||||
await self._client.aclose()
|
||||
self._client = None
|
||||
@@ -0,0 +1,276 @@
|
||||
"""OpenAI-compatible chat/completions adapter for text and vision."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
from typing import Any
|
||||
|
||||
import httpx
|
||||
|
||||
from sn_image_base.configs import is_valid_base_url
|
||||
from sn_image_base.exceptions import InvalidBaseUrlError, MissingApiKeyError
|
||||
from sn_image_base.utils.error_utils import (
|
||||
U1HttpBadResponseError,
|
||||
U1HttpNotFoundError,
|
||||
U1HttpResponseParseError,
|
||||
error_type_to_error_class,
|
||||
finish_reason_to_error_class,
|
||||
sanitize_base64_in_data,
|
||||
)
|
||||
from sn_image_base.utils.httpx_client import httpx_response_raise_for_status_code
|
||||
from sn_image_base.vlm.utils import image_to_data_url
|
||||
from sn_image_base.vlm.vlm_adapter import VlmAdapter
|
||||
|
||||
from .llm_adapter import LlmAdapter
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
DEFAULT_REQUEST_TIMEOUT = 600.0
|
||||
DEFAULT_MAX_COMPLETION_TOKENS = 8192
|
||||
|
||||
|
||||
class OpenAIChatAdapter(LlmAdapter, VlmAdapter):
|
||||
"""OpenAI-compatible ``/chat/completions`` adapter for text and vision."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
endpoint_url: str,
|
||||
api_key: str,
|
||||
model: str,
|
||||
*,
|
||||
timeout: float = DEFAULT_REQUEST_TIMEOUT,
|
||||
async_client: httpx.AsyncClient | None = None,
|
||||
reasoning_effort: str | None = None,
|
||||
) -> None:
|
||||
self._url = endpoint_url
|
||||
self._api_key = api_key
|
||||
self._default_model = model
|
||||
self._timeout = timeout
|
||||
self._reasoning_effort = reasoning_effort or None
|
||||
self._external_client = async_client
|
||||
self._client: httpx.AsyncClient | None = async_client
|
||||
logger.info(
|
||||
"OpenAIChatAdapter: endpoint=%s model=%s reasoning_effort=%s",
|
||||
self._url,
|
||||
self._default_model,
|
||||
self._reasoning_effort,
|
||||
)
|
||||
|
||||
def _get_client(self) -> httpx.AsyncClient:
|
||||
if self._client is None:
|
||||
self._client = httpx.AsyncClient(timeout=self._timeout)
|
||||
return self._client
|
||||
|
||||
@property
|
||||
def _headers(self) -> dict[str, str]:
|
||||
return {
|
||||
"Authorization": f"Bearer {self._api_key}",
|
||||
"Content-Type": "application/json",
|
||||
}
|
||||
|
||||
@staticmethod
|
||||
def _build_vision_content(
|
||||
user_prompt: str,
|
||||
images: list[str | bytes],
|
||||
) -> list[dict[str, Any]]:
|
||||
content: list[dict[str, Any]] = [{"type": "text", "text": user_prompt}]
|
||||
content.extend(
|
||||
{"type": "image_url", "image_url": {"url": image_to_data_url(img)}} for img in images
|
||||
)
|
||||
return content
|
||||
|
||||
def _build_payload(
|
||||
self,
|
||||
user_prompt: str,
|
||||
system_prompt: str,
|
||||
model: str,
|
||||
*,
|
||||
images: list[str | bytes] | None = None,
|
||||
max_completion_tokens: int | None = DEFAULT_MAX_COMPLETION_TOKENS,
|
||||
) -> dict[str, Any]:
|
||||
messages: list[dict[str, Any]] = []
|
||||
if system_prompt:
|
||||
messages.append({"role": "system", "content": system_prompt})
|
||||
|
||||
user_content: str | list[dict[str, Any]]
|
||||
if images:
|
||||
user_content = self._build_vision_content(user_prompt, images)
|
||||
else:
|
||||
user_content = user_prompt
|
||||
messages.append({"role": "user", "content": user_content})
|
||||
|
||||
payload: dict[str, Any] = {
|
||||
"model": model,
|
||||
"messages": messages,
|
||||
}
|
||||
if self._reasoning_effort:
|
||||
payload["reasoning_effort"] = self._reasoning_effort
|
||||
if max_completion_tokens:
|
||||
payload["max_completion_tokens"] = max_completion_tokens
|
||||
return payload
|
||||
|
||||
@staticmethod
|
||||
def _parse_response(data: dict[str, Any]) -> str:
|
||||
if "error" in data and (error := data["error"]):
|
||||
error_message = error.get("message")
|
||||
error_type = error.get("type")
|
||||
error_code = error.get("code")
|
||||
error_class, explanation = error_type_to_error_class(error_type)
|
||||
raise error_class(
|
||||
explanation,
|
||||
detail=f"chat/completions response has error. Error: {error_message}",
|
||||
code=error_code,
|
||||
)
|
||||
|
||||
choices = data.get("choices") or []
|
||||
if not choices:
|
||||
sanitized_data = sanitize_base64_in_data(data)
|
||||
dumped = json.dumps(sanitized_data, ensure_ascii=False)
|
||||
raise U1HttpBadResponseError(
|
||||
detail=f"chat/completions response has no choices. Response: {dumped}",
|
||||
)
|
||||
|
||||
contents: list[str] = []
|
||||
finish_reason: str | None = None
|
||||
for choice in choices:
|
||||
msg = choice.get("message", {})
|
||||
finish_reason = choice.get("finish_reason") or finish_reason
|
||||
content_val = msg.get("content")
|
||||
if isinstance(content_val, str):
|
||||
contents.append(content_val)
|
||||
elif isinstance(content_val, list):
|
||||
parts: list[str] = []
|
||||
for block in content_val:
|
||||
if isinstance(block, dict) and block.get("type") == "text":
|
||||
text = block.get("text")
|
||||
if isinstance(text, str):
|
||||
parts.append(text)
|
||||
contents.append("".join(parts))
|
||||
|
||||
final_content = "".join(contents)
|
||||
if final_content:
|
||||
return final_content
|
||||
|
||||
sanitized_data = sanitize_base64_in_data(data)
|
||||
dumped = json.dumps(sanitized_data, ensure_ascii=False)
|
||||
detail_msg = ""
|
||||
if finish_reason:
|
||||
detail_msg += f"\n^ Finish reason: {finish_reason}"
|
||||
detail_msg += f"\n^ Response: {dumped}"
|
||||
if finish_reason == "stop":
|
||||
raise U1HttpBadResponseError(
|
||||
"chat/completions response with empty content.",
|
||||
detail=detail_msg,
|
||||
)
|
||||
if finish_reason:
|
||||
error_class, explanation = finish_reason_to_error_class(finish_reason)
|
||||
raise error_class(explanation, detail=detail_msg)
|
||||
raise U1HttpBadResponseError(
|
||||
"chat/completions response has no content. No finish reason provided.",
|
||||
detail=detail_msg,
|
||||
)
|
||||
|
||||
async def _post_payload(self, payload: dict[str, Any], model: str) -> str:
|
||||
resp = await self._get_client().post(self._url, json=payload, headers=self._headers)
|
||||
try:
|
||||
httpx_response_raise_for_status_code(resp)
|
||||
data = resp.json()
|
||||
except U1HttpNotFoundError as exc:
|
||||
raise U1HttpNotFoundError(
|
||||
detail=f"{exc.detail} model={model!r}",
|
||||
code=resp.status_code,
|
||||
) from exc
|
||||
except ValueError as exc:
|
||||
raise U1HttpResponseParseError(
|
||||
detail=f"Failed to parse HTTP response. {resp.request.url}. Response content: {resp.content}",
|
||||
code=resp.status_code,
|
||||
) from exc
|
||||
return self._parse_response(data)
|
||||
|
||||
async def text_completion(
|
||||
self,
|
||||
user_prompt: str,
|
||||
system_prompt: str = "",
|
||||
model: str | None = None,
|
||||
) -> str:
|
||||
resolved_model = model or self._default_model
|
||||
payload = self._build_payload(user_prompt, system_prompt, resolved_model)
|
||||
return await self._post_payload(payload, resolved_model)
|
||||
|
||||
async def vision_completion(
|
||||
self,
|
||||
user_prompt: str,
|
||||
images: list[str | bytes],
|
||||
system_prompt: str = "",
|
||||
model: str | None = None,
|
||||
) -> str:
|
||||
resolved_model = model or self._default_model
|
||||
payload = self._build_payload(
|
||||
user_prompt,
|
||||
system_prompt,
|
||||
resolved_model,
|
||||
images=images,
|
||||
)
|
||||
return await self._post_payload(payload, resolved_model)
|
||||
|
||||
async def aclose(self) -> None:
|
||||
if self._external_client is None and self._client is not None:
|
||||
await self._client.aclose()
|
||||
self._client = None
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import argparse
|
||||
import asyncio
|
||||
|
||||
from sn_image_base.configs import global_configs
|
||||
|
||||
parser = argparse.ArgumentParser(description="Async OpenAI-compatible chat adapter.")
|
||||
parser.add_argument("--prompt", default=None, help="Prompt to use for the model")
|
||||
parser.add_argument("--system-prompt", default=None, help="System prompt to use")
|
||||
parser.add_argument("--image", default=os.environ.get("IMAGE_PATH"), help="Optional image path")
|
||||
args = parser.parse_args()
|
||||
|
||||
async def main() -> None:
|
||||
prompt = args.prompt or "Write a poem about the topic: 'Hello world'"
|
||||
base_url = global_configs.SN_CHAT_BASE_URL
|
||||
if not base_url:
|
||||
raise InvalidBaseUrlError(
|
||||
f"No base URL provided for chat runtime. {global_configs.get_env_var_help('SN_CHAT_BASE_URL')}"
|
||||
)
|
||||
if not is_valid_base_url(base_url):
|
||||
raise InvalidBaseUrlError(
|
||||
f"Invalid base URL for chat runtime: {base_url}. {global_configs.get_env_var_help('SN_CHAT_BASE_URL')}"
|
||||
)
|
||||
endpoint_url = f"{base_url.rstrip('/')}/chat/completions"
|
||||
api_key = global_configs.SN_CHAT_API_KEY
|
||||
if not api_key:
|
||||
raise MissingApiKeyError(
|
||||
f"No API key provided for chat runtime. {global_configs.get_env_var_help('SN_CHAT_API_KEY')}"
|
||||
)
|
||||
model = global_configs.SN_TEXT_MODEL
|
||||
|
||||
adapter = OpenAIChatAdapter(
|
||||
endpoint_url=endpoint_url,
|
||||
api_key=api_key,
|
||||
model=model,
|
||||
)
|
||||
try:
|
||||
if args.image:
|
||||
result = await adapter.vision_completion(
|
||||
user_prompt=prompt,
|
||||
images=[args.image],
|
||||
system_prompt=args.system_prompt or "",
|
||||
)
|
||||
else:
|
||||
result = await adapter.text_completion(
|
||||
user_prompt=prompt,
|
||||
system_prompt=args.system_prompt or "",
|
||||
)
|
||||
print(result)
|
||||
finally:
|
||||
await adapter.aclose()
|
||||
|
||||
asyncio.run(main())
|
||||
51
sn-image-base/scripts/sn_image_base/llm/llm_adapter.py
Normal file
51
sn-image-base/scripts/sn_image_base/llm/llm_adapter.py
Normal file
@@ -0,0 +1,51 @@
|
||||
"""Abstract base class for LLM (Language Model) adapters."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from abc import ABC, abstractmethod
|
||||
|
||||
|
||||
class LlmAdapter(ABC):
|
||||
"""Uniform async interface for a single Language Model backend.
|
||||
|
||||
Each concrete adapter wraps one LLM endpoint + model combination and
|
||||
exposes a single :meth:`text_completion` coroutine. Synchronous
|
||||
calling is intentionally **not** supported; callers must run inside an
|
||||
asyncio event loop.
|
||||
|
||||
**Client ownership contract** — when a shared
|
||||
:class:`httpx.AsyncClient` is supplied at construction time the adapter
|
||||
*reuses* it and must **not** close it; the caller retains full ownership
|
||||
of the client's lifecycle. When no external client is provided the
|
||||
adapter creates and owns an internal client and must close it in
|
||||
:meth:`aclose`.
|
||||
"""
|
||||
|
||||
@abstractmethod
|
||||
async def text_completion(
|
||||
self,
|
||||
user_prompt: str,
|
||||
system_prompt: str = "",
|
||||
model: str | None = None,
|
||||
) -> str:
|
||||
"""Send a text-only prompt to the model and return the reply.
|
||||
|
||||
Args:
|
||||
user_prompt: User-facing text instruction.
|
||||
system_prompt: System-level instruction prepended to the
|
||||
conversation. Defaults to ''.
|
||||
model: Model name to use. If None, uses the default set at
|
||||
initialization.
|
||||
|
||||
Returns:
|
||||
str: Raw text response from the model.
|
||||
"""
|
||||
|
||||
@abstractmethod
|
||||
async def aclose(self) -> None:
|
||||
"""Release async resources owned by this adapter.
|
||||
|
||||
Must be called when the adapter is no longer needed. Adapters that
|
||||
were given an external shared client must implement this as a no-op;
|
||||
adapters that created their own internal client must close it here.
|
||||
"""
|
||||
Reference in New Issue
Block a user