This commit is contained in:
xuhd committed 2026-05-16 15:51:49 +08:00
commit 48ddbf0412
423 files changed
+117115 -3214

No files matched your search

@@ -0,0 +1,78 @@
"""Middleware that injects the admin-configured prompt prefix at model-call time.
The prompt-prefix feature lets an admin prepend a fixed instruction to the
user's question before it reaches the LLM. To keep the prefix invisible in the
chat UI, the Gateway does **not** concatenate it into the persisted
``HumanMessage.content``; instead it tags the first human message with
``additional_kwargs["prompt_prefix"]`` (see ``app.gateway.services``).
This middleware reads that tag and prepends the prefix to the message content
*only in the request handed to the model*. The persisted state — and therefore
what the frontend renders — keeps the clean, prefix-free content.
"""
from collections.abc import Awaitable, Callable
from typing import override
from langchain.agents import AgentState
from langchain.agents.middleware import AgentMiddleware
from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
from langchain_core.messages import HumanMessage
PROMPT_PREFIX_KEY = "prompt_prefix"
class PromptPrefixMiddleware(AgentMiddleware[AgentState]):
"""Prepends ``additional_kwargs['prompt_prefix']`` to human messages for the LLM only.
The transformation is applied to the model request copy, never written back
to state, so it is naturally idempotent across the multiple model calls of a
single turn.
"""
@staticmethod
def _apply(messages: list) -> list | None:
"""Return a patched copy with prefixes prepended, or None if nothing to do."""
patched: list | None = None
for idx, msg in enumerate(messages):
if not isinstance(msg, HumanMessage):
continue
raw_prefix = (getattr(msg, "additional_kwargs", None) or {}).get(PROMPT_PREFIX_KEY)
prefix = str(raw_prefix).strip() if raw_prefix else ""
if not prefix:
continue
content = msg.content
if isinstance(content, str):
new_content: object = f"{prefix}\n\n{content}".strip("\n")
elif isinstance(content, list):
# Structured content (text blocks + media): prepend a text block.
new_content = [{"type": "text", "text": prefix}, *content]
else:
new_content = f"{prefix}\n\n{content!s}"
if patched is None:
patched = list(messages)
# model_copy preserves the message id so streaming dedup is unaffected.
patched[idx] = msg.model_copy(update={"content": new_content})
return patched
@override
def wrap_model_call(
self,
request: ModelRequest,
handler: Callable[[ModelRequest], ModelResponse],
) -> ModelCallResult:
patched = self._apply(request.messages)
if patched is not None:
request = request.override(messages=patched)
return handler(request)
@override
async def awrap_model_call(
self,
request: ModelRequest,
handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
) -> ModelCallResult:
patched = self._apply(request.messages)
if patched is not None:
request = request.override(messages=patched)
return await handler(request)
@@ -72,7 +72,7 @@ class SetupAgentApprovalMiddleware(AgentMiddleware[SetupAgentApprovalMiddlewareS
approval_msg = ToolMessage(
id=self._stable_message_id(tool_call_id, args),
content="请确认是否创建/更新这个专家配置。",
content="请确认是否创建/更新这个智能体配置。",
tool_call_id=tool_call_id,
name="setup_agent",
additional_kwargs={"setup_agent_approval": payload},