修改了nanobot，往Hermes agent的风格走，进度1/3
This commit is contained in:
6
app-instance/backend/beaver/skills/assembler/__init__.py
Normal file
6
app-instance/backend/beaver/skills/assembler/__init__.py
Normal file
@ -0,0 +1,6 @@
|
||||
"""Skill assembly for Beaver."""
|
||||
|
||||
from .embedding_retriever import SkillEmbeddingRetriever
|
||||
from .task_assembler import SkillAssemblyResult, SkillAssembler
|
||||
|
||||
__all__ = ["SkillAssemblyResult", "SkillAssembler", "SkillEmbeddingRetriever"]
|
||||
@ -0,0 +1,188 @@
|
||||
"""Embedding-based skill candidate retrieval.
|
||||
|
||||
当前实现使用 OpenAI-compatible `/v1/embeddings` 接口调用
|
||||
阿里云百炼 `text-embedding-v4` 做最小语义召回：
|
||||
1. 复用当前 provider 的 `api_key/api_base`
|
||||
2. 先用 embedding 相似度召回一小批候选
|
||||
3. 再交给上层 LLM selector 做最终技能选择
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import math
|
||||
import os
|
||||
import json
|
||||
from urllib import request
|
||||
from typing import Any
|
||||
|
||||
|
||||
class SkillEmbeddingRetriever:
|
||||
"""用 OpenAI-compatible embeddings API 为 skill 选择做候选召回。"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
api_key_env: str = "OPENAI_API_KEY",
|
||||
api_base_env: str = "OPENAI_API_BASE",
|
||||
model: str = "text-embedding-v4",
|
||||
timeout_seconds: float = 20.0,
|
||||
) -> None:
|
||||
self.api_key_env = api_key_env
|
||||
self.api_base_env = api_base_env
|
||||
self.model = model
|
||||
self.timeout_seconds = timeout_seconds
|
||||
|
||||
async def retrieve(
|
||||
self,
|
||||
*,
|
||||
query: str,
|
||||
candidates: list[dict[str, str]],
|
||||
top_k: int = 12,
|
||||
api_key: str | None = None,
|
||||
api_base: str | None = None,
|
||||
model: str | None = None,
|
||||
) -> list[dict[str, str]]:
|
||||
"""按 embedding 相似度召回 top-k 候选。
|
||||
|
||||
如果没有可用的 API Key / base URL，或者 embedding 调用失败，
|
||||
当前阶段先退回到“全部候选交给 LLM selector”。
|
||||
"""
|
||||
|
||||
if not candidates:
|
||||
return []
|
||||
|
||||
resolved_api_key = api_key or os.getenv(self.api_key_env)
|
||||
resolved_api_base = api_base or os.getenv(self.api_base_env)
|
||||
if not resolved_api_key or not resolved_api_base:
|
||||
return candidates
|
||||
|
||||
try:
|
||||
query_embedding = await self._embed_texts(
|
||||
api_key=resolved_api_key,
|
||||
api_base=resolved_api_base,
|
||||
texts=[query],
|
||||
model=model or self.model,
|
||||
)
|
||||
candidate_texts = [self._candidate_text(item) for item in candidates]
|
||||
candidate_embeddings = await self._embed_texts(
|
||||
api_key=resolved_api_key,
|
||||
api_base=resolved_api_base,
|
||||
texts=candidate_texts,
|
||||
model=model or self.model,
|
||||
)
|
||||
except Exception:
|
||||
return candidates
|
||||
|
||||
if not query_embedding or not query_embedding[0] or len(candidate_embeddings) != len(candidates):
|
||||
return candidates
|
||||
|
||||
query_vector = query_embedding[0]
|
||||
scored: list[tuple[float, dict[str, str]]] = []
|
||||
for candidate, vector in zip(candidates, candidate_embeddings, strict=False):
|
||||
if not vector:
|
||||
continue
|
||||
scored.append((self._cosine_similarity(query_vector, vector), candidate))
|
||||
|
||||
scored.sort(key=lambda item: item[0], reverse=True)
|
||||
return [item[1] for item in scored[:top_k]]
|
||||
|
||||
async def _embed_texts(
|
||||
self,
|
||||
*,
|
||||
api_key: str,
|
||||
api_base: str,
|
||||
texts: list[str],
|
||||
model: str,
|
||||
) -> list[list[float]]:
|
||||
"""调用 OpenAI-compatible embeddings 接口。
|
||||
|
||||
当前对齐的是你们实际在用的网关配置：
|
||||
- `POST {api_base}/embeddings`
|
||||
- `model=text-embedding-v4`
|
||||
- `encoding_format=float`
|
||||
"""
|
||||
|
||||
all_vectors: list[list[float]] = []
|
||||
endpoint = self._normalize_embeddings_endpoint(api_base)
|
||||
for start in range(0, len(texts), 10):
|
||||
batch = texts[start:start + 10]
|
||||
payload = await self._post_embeddings(
|
||||
endpoint=endpoint,
|
||||
api_key=api_key,
|
||||
model=model,
|
||||
texts=batch,
|
||||
)
|
||||
embeddings = payload.get("data") or []
|
||||
embeddings = sorted(embeddings, key=lambda item: item.get("index", 0))
|
||||
all_vectors.extend([list(item.get("embedding") or []) for item in embeddings])
|
||||
return all_vectors
|
||||
|
||||
async def _post_embeddings(
|
||||
self,
|
||||
*,
|
||||
endpoint: str,
|
||||
api_key: str,
|
||||
model: str,
|
||||
texts: list[str],
|
||||
) -> dict[str, Any]:
|
||||
return await asyncio.to_thread(
|
||||
self._post_embeddings_sync,
|
||||
endpoint=endpoint,
|
||||
api_key=api_key,
|
||||
model=model,
|
||||
texts=texts,
|
||||
)
|
||||
|
||||
def _post_embeddings_sync(
|
||||
self,
|
||||
*,
|
||||
endpoint: str,
|
||||
api_key: str,
|
||||
model: str,
|
||||
texts: list[str],
|
||||
) -> dict[str, Any]:
|
||||
body = json.dumps(
|
||||
{
|
||||
"model": model,
|
||||
"input": texts if len(texts) > 1 else texts[0],
|
||||
"encoding_format": "float",
|
||||
}
|
||||
).encode("utf-8")
|
||||
req = request.Request(
|
||||
endpoint,
|
||||
data=body,
|
||||
headers={
|
||||
"Authorization": f"Bearer {api_key}",
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
method="POST",
|
||||
)
|
||||
with request.urlopen(req, timeout=self.timeout_seconds) as response:
|
||||
return json.loads(response.read().decode("utf-8"))
|
||||
|
||||
@staticmethod
|
||||
def _candidate_text(candidate: dict[str, str]) -> str:
|
||||
name = (candidate.get("name") or "").strip()
|
||||
description = (candidate.get("description") or "").strip()
|
||||
return f"{name}\n{description}".strip()
|
||||
|
||||
@staticmethod
|
||||
def _normalize_embeddings_endpoint(api_base: str) -> str:
|
||||
base = api_base.rstrip("/")
|
||||
if base.endswith("/embeddings"):
|
||||
return base
|
||||
if base.endswith("/v1"):
|
||||
return f"{base}/embeddings"
|
||||
return f"{base}/v1/embeddings"
|
||||
|
||||
@staticmethod
|
||||
def _cosine_similarity(left: list[float], right: list[float]) -> float:
|
||||
if not left or not right or len(left) != len(right):
|
||||
return -1.0
|
||||
dot = sum(a * b for a, b in zip(left, right, strict=False))
|
||||
left_norm = math.sqrt(sum(a * a for a in left))
|
||||
right_norm = math.sqrt(sum(b * b for b in right))
|
||||
if left_norm == 0 or right_norm == 0:
|
||||
return -1.0
|
||||
return dot / (left_norm * right_norm)
|
||||
168
app-instance/backend/beaver/skills/assembler/task_assembler.py
Normal file
168
app-instance/backend/beaver/skills/assembler/task_assembler.py
Normal file
@ -0,0 +1,168 @@
|
||||
"""LLM-driven skill assembler.
|
||||
|
||||
这层现在不再自己做规则打分，而是直接把：
|
||||
1. task description
|
||||
2. embedding 召回后的候选 skill 摘要
|
||||
|
||||
交给一个模型来决定本轮要激活哪些 skill。
|
||||
|
||||
当前目标非常克制：
|
||||
- 输入尽量简单
|
||||
- 输出只要 skill 名称
|
||||
- 没有命中就返回空 skills
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import dataclass, field
|
||||
import json
|
||||
from typing import Any
|
||||
|
||||
from beaver.engine.context import SkillContext
|
||||
from beaver.engine.providers.base import LLMProvider
|
||||
from beaver.engine.providers.runtime import ProviderRuntime
|
||||
from beaver.skills.catalog.loader import SkillsLoader
|
||||
from beaver.skills.catalog.utils import strip_frontmatter
|
||||
from .embedding_retriever import SkillEmbeddingRetriever
|
||||
|
||||
|
||||
@dataclass(slots=True)
|
||||
class SkillAssemblyResult:
|
||||
"""一次装配后真正要注入当前 run 的 skills。"""
|
||||
|
||||
activated_skills: list[SkillContext] = field(default_factory=list)
|
||||
|
||||
|
||||
class SkillAssembler:
|
||||
"""用 LLM 根据 task description 选择当前 run 的 skills。"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
loader: SkillsLoader,
|
||||
retriever: SkillEmbeddingRetriever | None = None,
|
||||
) -> None:
|
||||
self.loader = loader
|
||||
self.retriever = retriever or SkillEmbeddingRetriever()
|
||||
|
||||
async def assemble(
|
||||
self,
|
||||
*,
|
||||
task_description: str,
|
||||
provider: LLMProvider,
|
||||
model: str,
|
||||
embedding_runtime: ProviderRuntime | None = None,
|
||||
top_k: int = 12,
|
||||
) -> SkillAssemblyResult:
|
||||
candidates = self.loader.build_selection_candidates()
|
||||
if not candidates:
|
||||
return SkillAssemblyResult()
|
||||
candidates = await self.retriever.retrieve(
|
||||
query=task_description,
|
||||
candidates=candidates,
|
||||
top_k=top_k,
|
||||
api_key=embedding_runtime.api_key if embedding_runtime is not None else None,
|
||||
api_base=embedding_runtime.api_base if embedding_runtime is not None else None,
|
||||
model=embedding_runtime.model if embedding_runtime is not None else None,
|
||||
)
|
||||
if not candidates:
|
||||
return SkillAssemblyResult()
|
||||
|
||||
selected_names = await self._select_skill_names(
|
||||
task_description=task_description,
|
||||
candidates=candidates,
|
||||
provider=provider,
|
||||
model=model,
|
||||
)
|
||||
if not selected_names:
|
||||
return SkillAssemblyResult()
|
||||
|
||||
activated_skills: list[SkillContext] = []
|
||||
for name in selected_names:
|
||||
raw_content = self.loader.load_skill(name)
|
||||
content = strip_frontmatter(raw_content).strip() if raw_content else ""
|
||||
if not content:
|
||||
continue
|
||||
activated_skills.append(SkillContext(name=name, content=content))
|
||||
|
||||
return SkillAssemblyResult(activated_skills=activated_skills)
|
||||
|
||||
async def _select_skill_names(
|
||||
self,
|
||||
*,
|
||||
task_description: str,
|
||||
candidates: list[dict[str, str]],
|
||||
provider: LLMProvider,
|
||||
model: str,
|
||||
) -> list[str]:
|
||||
candidate_summary = self._render_candidates(candidates)
|
||||
candidate_names = {item["name"] for item in candidates}
|
||||
messages = [
|
||||
{
|
||||
"role": "system",
|
||||
"content": (
|
||||
"You select Beaver skills for a single run. "
|
||||
"Given a task description and candidate skill summaries, "
|
||||
"return only a JSON array of skill names to activate. "
|
||||
"Do not invent names. If nothing matches, return []."
|
||||
),
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": (
|
||||
f"Task description:\n{task_description}\n\n"
|
||||
f"Candidate skills:\n{candidate_summary}\n\n"
|
||||
"Return only JSON, for example: [\"skill-a\", \"skill-b\"]"
|
||||
),
|
||||
},
|
||||
]
|
||||
response = await provider.chat(
|
||||
messages=messages,
|
||||
tools=None,
|
||||
model=model,
|
||||
max_tokens=512,
|
||||
temperature=0,
|
||||
)
|
||||
if response.finish_reason == "error" or not response.content:
|
||||
return []
|
||||
|
||||
parsed = self._parse_selected_names(response.content)
|
||||
if not parsed:
|
||||
return []
|
||||
|
||||
# 只保留当前候选集中真实存在的 skill 名称，并维持模型输出顺序。
|
||||
filtered: list[str] = []
|
||||
for name in parsed:
|
||||
if name in candidate_names and name not in filtered:
|
||||
filtered.append(name)
|
||||
return filtered
|
||||
|
||||
@staticmethod
|
||||
def _render_candidates(candidates: list[dict[str, str]]) -> str:
|
||||
lines: list[str] = []
|
||||
for item in candidates:
|
||||
lines.append(f"- {item['name']}: {item['description']}")
|
||||
return "\n".join(lines)
|
||||
|
||||
@staticmethod
|
||||
def _parse_selected_names(content: str) -> list[str]:
|
||||
cleaned = content.strip()
|
||||
if cleaned.startswith("```"):
|
||||
lines = cleaned.splitlines()
|
||||
if len(lines) >= 3 and lines[0].startswith("```") and lines[-1].startswith("```"):
|
||||
cleaned = "\n".join(lines[1:-1]).strip()
|
||||
|
||||
try:
|
||||
payload: Any = json.loads(cleaned)
|
||||
except json.JSONDecodeError:
|
||||
return []
|
||||
|
||||
if isinstance(payload, dict):
|
||||
for key in ("skills", "selected_skills", "activated_skills", "selected"):
|
||||
value = payload.get(key)
|
||||
if isinstance(value, list):
|
||||
payload = value
|
||||
break
|
||||
|
||||
if not isinstance(payload, list):
|
||||
return []
|
||||
return [item.strip() for item in payload if isinstance(item, str) and item.strip()]
|
||||
Reference in New Issue
Block a user