This commit is contained in:
fengxiangdebaozi committed 2026-06-03 18:54:26 +08:00
commit 8f6e08eb2f
102 files changed
+6816 -6016

No files matched your search

@@ -1,8 +1,9 @@
"""素材收集专家节点:关键词提取 → 并发搜索 → 相关度排序。"""
"""素材收集专家节点:关键词提取 → 并发搜索(含笔记) → 相关度排序。"""
from __future__ import annotations
import asyncio
import logging
import os
import re
import time
import uuid
@@ -21,6 +22,10 @@ from ._utils import get_ai_writing_config, llm_json, safe_get_stream_writer
logger = logging.getLogger(__name__)
_NOTEBOOK_SEARCH_PATH = "/api/v1/source/mentioned-documents/search"
_NOTEBOOK_MAX_KEYWORDS = 3 # 每次写作最多对笔记发起的检索轮数
_NOTEBOOK_TOP_K = 5 # 每次检索取回的 chunk 数
# 素材总量上限边界:config.yaml 的 ai_writing.max_materials,非法值回退默认 50
_DEFAULT_MAX_MATERIALS = 50
_MIN_MAX_MATERIALS = 1
@@ -95,6 +100,67 @@ def _notice(title: str, detail: str, raw: str = "") -> dict:
return event
async def _search_notebook(
keywords: list[str],
search_space_id: int,
source_ids: list[int],
token: str,
) -> list[Material]:
"""从笔记 API 按关键词检索内容,返回 source_type='notebook' 的 Material 列表。"""
import httpx # 懒导入:仅在笔记模式下才需要
base = (os.getenv("DEER_FLOW_NOTEBOOK_API_BASE") or "").rstrip("/")
if not base:
logger.warning("DEER_FLOW_NOTEBOOK_API_BASE not set, skipping notebook search")
return []
results: list[Material] = []
mat_idx = 0
try:
async with httpx.AsyncClient(timeout=20) as client:
for kw in keywords[:_NOTEBOOK_MAX_KEYWORDS]:
try:
resp = await client.post(
f"{base}{_NOTEBOOK_SEARCH_PATH}",
headers={"Authorization": f"Bearer {token}"},
json={
"query": kw,
"source_ids": source_ids,
"search_space_id": search_space_id,
"top_k": _NOTEBOOK_TOP_K,
},
)
resp.raise_for_status()
data = resp.json()
# 解析 documents 列表,每个 chunk 作为独立素材
documents: list[dict] = data.get("documents") if isinstance(data, dict) else []
for doc in documents or []:
doc_title: str = doc.get("title") or "笔记文档"
doc_url: str = doc.get("url") or ""
for chunk in doc.get("chunks") or []:
chunk_content: str = (chunk.get("content") or "").strip()
if not chunk_content:
continue
mat_idx += 1
raw_score = chunk.get("score")
chunk_score = float(raw_score) if raw_score is not None else max(0.5, 0.95 - (mat_idx - 1) * 0.03)
results.append(Material(
id=f"nb{mat_idx:03d}",
title=doc_title,
content=chunk_content[:500],
source=f"笔记·{doc_title}",
source_type="notebook",
relevance_score=chunk_score,
url=doc_url,
))
except Exception as e:
logger.warning("notebook search kw=%r failed: %s", kw, e)
except Exception as e:
logger.warning("notebook search client error: %s", e)
return results
async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict:
intent = state["user_intent"]
article_type = state.get("article_type", "analysis")
@@ -181,6 +247,22 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
if extra_keywords:
keywords = list(dict.fromkeys(keywords + extra_keywords))
# Step 1.5: 笔记检索(如果传入了 notebook_search_space_id)
notebook_search_space_id: int | None = state.get("notebook_search_space_id")
notebook_materials: list[Material] = []
if notebook_search_space_id:
notebook_source_ids: list[int] = state.get("notebook_source_ids") or []
notebook_token: str = cfg.get("notebook_token") or ""
_emit("正在检索笔记内容…")
notebook_materials = await _search_notebook(
keywords[:_NOTEBOOK_MAX_KEYWORDS],
notebook_search_space_id,
notebook_source_ids,
notebook_token,
)
if notebook_materials:
_emit(f"从笔记中获取到 {len(notebook_materials)} 条素材")
# Step 2: 搜索
# 技能检索:技能不是素材,是"检索策略 / 视角"。流程:
# a. SkillPicker 选出按相关度排序的 skills(推 list_skills + select_skills 事件)
@@ -330,10 +412,17 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
existing_pkg = state.get("material_package")
if is_skill_search:
# 技能编排循环已经完成素材累积,直接用 raw_materials_pre 当 raw_materials
# 笔记素材追加在技能素材之后
raw_materials: list[Material] = raw_materials_pre # type: ignore[name-defined]
raw_materials.extend(notebook_materials)
else:
seen_urls: set[str] = set()
raw_materials = []
# 笔记素材优先排在最前面,赋予高相关度
raw_materials = list(notebook_materials)
for m in notebook_materials:
if m.get("url"):
seen_urls.add(m["url"])
preserve_existing = existing_pkg and (is_supplement or is_append_only)
if preserve_existing:
for m in existing_pkg.get("materials", []):
@@ -143,6 +143,9 @@ class AIWritingState(TypedDict):
# 经 checkpoint 持久化 —— interrupt/resume 是新 invocation、其 config 不带原
# model_name,存进 state 才能让 resume 后才运行的节点拿到用户模型。
model_name: str
# 笔记来源(可选):传入后 researcher 阶段会额外从笔记检索素材
notebook_search_space_id: Optional[int]
notebook_source_ids: Optional[List[int]]
# Agent A 输出
material_package: Optional[MaterialPackage]
@@ -285,18 +285,28 @@ def _build_middlewares(
if resolved_app_config.token_usage.enabled:
middlewares.append(TokenUsageMiddleware())
# Add TitleMiddleware
middlewares.append(TitleMiddleware(app_config=resolved_app_config))
# Add TitleMiddleware — but skip it for scheduled runs. Each scheduled run
# executes on a throwaway thread whose title is never shown to a user, and
# generating one would fire an extra LLM call on the *default* model
# (config.title.model_name is usually unset → models[0]), which is not the
# model the task selected. Skipping it keeps scheduled/HTML runs on exactly
# the chosen model.
if not cfg.get("is_scheduled_run", False):
middlewares.append(TitleMiddleware(app_config=resolved_app_config))
# Add MemoryMiddleware (after TitleMiddleware) — skip in bootstrap mode (agent creation flow)
memory_injection_enabled = cfg.get("memory_injection_enabled", True)
if not is_bootstrap and not is_writing_setup:
middlewares.append(MemoryMiddleware(agent_name=agent_name, memory_config=resolved_app_config.memory, injection_enabled=memory_injection_enabled))
# Add ViewImageMiddleware only if the current model supports vision.
# Use the resolved runtime model_name from make_lead_agent to avoid stale config values.
model_config = resolved_app_config.get_model_config(model_name) if model_name else None
if model_config is not None and model_config.supports_vision:
# Add ViewImageMiddleware only for the NATIVE-vision path (main model
# ingests raw image data). In delegated mode the view_image tool already
# returns a text description and stores nothing in `viewed_images`, so the
# middleware would otherwise inject a spurious "No images have been viewed."
# message. Use the resolved runtime model_name to avoid stale config values.
from deerflow.models.vision import is_delegated_vision, vision_enabled
if vision_enabled(model_name, app_config=resolved_app_config) and not is_delegated_vision(app_config=resolved_app_config):
middlewares.append(ViewImageMiddleware())
# Add DeferredToolFilterMiddleware to hide deferred tool schemas from model binding
@@ -18,6 +18,12 @@ logger = logging.getLogger(__name__)
_OUTLINE_PREVIEW_LINES = 5
# Image files are understood via the `view_image` tool (which routes to the
# configured vision model), never by reading their bytes. Listing them with the
# generic read_file/grep guidance made the agent try to parse image binaries
# with bash/python, which fails and wastes turns.
_IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".webp", ".gif", ".bmp"}
def _extract_outline_for_file(file_path: Path) -> tuple[list[dict], list[str]]:
"""Return the document outline and fallback preview for *file_path*.
@@ -88,6 +94,16 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
size_str = f"{size_kb:.1f} KB" if size_kb < 1024 else f"{size_kb / 1024:.1f} MB"
lines.append(f"- {file['filename']} ({size_str})")
lines.append(f" Path: {file['path']}")
# Images must be understood through the vision model, not by reading bytes.
extension = str(file.get("extension") or Path(file["filename"]).suffix).lower()
if extension in _IMAGE_EXTENSIONS:
lines.append(f" This is an image. To understand it, call the `view_image` tool with image_path='{file['path']}' (optionally pass a `question`).")
lines.append(" The configured vision model will analyze it and return a description — trust that result.")
lines.append(" Do NOT read, open, or parse the image with read_file, grep, bash, python, PIL, or any byte-level inspection; those cannot interpret image content.")
lines.append("")
return
outline = file.get("outline") or []
if outline:
truncated = outline[-1].get("truncated", False)
@@ -139,6 +139,12 @@ class ViewImageMiddleware(AgentMiddleware[ViewImageMiddlewareState]):
if not messages:
return False
# Nothing to inject when no image data was captured (e.g. delegated
# recognition handled the image in the tool and stored no base64).
# Avoids injecting a useless "No images have been viewed." message.
if not state.get("viewed_images"):
return False
# Get the last assistant message
last_assistant_msg = self._get_last_assistant_message(messages)
if not last_assistant_msg: