feat(openwebui-skills-manager): enhance auto-discovery and structural refactoring
- Enable default overwrite installation policy for overlapping skills - Support deep recursive GitHub trees discovery mechanism to resolve #58 - Refactor internal architecture to fully decouple stateless helper logic - READMEs and docs synced (v0.3.0)
This commit is contained in:
fujie
206
plugins/debug/byok-infinite-session-research/analysis.md
Normal file
206
plugins/debug/byok-infinite-session-research/analysis.md
Normal file
@@ -0,0 +1,206 @@
|
||||
# BYOK模式与Infinite Session(自动上下文压缩)兼容性研究
|
||||
|
||||
**日期**: 2026-03-08
|
||||
**研究范围**: Copilot SDK v0.1.30 + OpenWebUI Extensions Pipe v0.10.0
|
||||
|
||||
## 研究问题
|
||||
在BYOK (Bring Your Own Key) 模式下，是否应该支持自动上下文压缩(Infinite Sessions)?
|
||||
用户报告：BYOK模式本不应该触发压缩，但当模型名称与Copilot内置模型一致时，意外地支持了压缩。
|
||||
|
||||
---
|
||||
|
||||
## 核心发现
|
||||
|
||||
### 1. SDK层面（copilot-sdk/python/copilot/types.py）
|
||||
|
||||
**InfiniteSessionConfig 定义** (line 453-470):
|
||||
```python
|
||||
class InfiniteSessionConfig(TypedDict, total=False):
|
||||
"""
|
||||
Configuration for infinite sessions with automatic context compaction
|
||||
and workspace persistence.
|
||||
"""
|
||||
enabled: bool
|
||||
background_compaction_threshold: float # 0.0-1.0, default: 0.80
|
||||
buffer_exhaustion_threshold: float # 0.0-1.0, default: 0.95
|
||||
```
|
||||
|
||||
**SessionConfig结构** (line 475+):
|
||||
- `provider: ProviderConfig` - 用于BYOK配置
|
||||
- `infinite_sessions: InfiniteSessionConfig` - 上下文压缩配置
|
||||
- **关键**: 这两个配置是**完全独立的**，没有相互依赖关系
|
||||
|
||||
### 2. OpenWebUI Pipe层面（github_copilot_sdk.py）
|
||||
|
||||
**Infinite Session初始化** (line 5063-5069):
|
||||
```python
|
||||
infinite_session_config = None
|
||||
if self.valves.INFINITE_SESSION: # 默认值: True
|
||||
infinite_session_config = InfiniteSessionConfig(
|
||||
enabled=True,
|
||||
background_compaction_threshold=self.valves.COMPACTION_THRESHOLD,
|
||||
buffer_exhaustion_threshold=self.valves.BUFFER_THRESHOLD,
|
||||
)
|
||||
```
|
||||
|
||||
**关键问题**:
|
||||
- ✗ 没有任何条件检查 `is_byok_model`
|
||||
- ✗ 无论使用官方模型还是BYOK模型，都会应用相同的infinite session配置
|
||||
- ✓ 回对比，reasoning_effort被正确地在BYOK模式下禁用（line 6329-6331）
|
||||
|
||||
### 3. 模型识别逻辑（line 6199+）
|
||||
|
||||
```python
|
||||
if m_info and "source" in m_info:
|
||||
is_byok_model = m_info["source"] == "byok"
|
||||
else:
|
||||
is_byok_model = not has_multiplier and byok_active
|
||||
```
|
||||
|
||||
BYOK模型识别基于:
|
||||
1. 模型元数据中的 `source` 字段
|
||||
2. 或者根据是否有乘数标签 (如 "4x", "0.5x") 和globally active的BYOK配置
|
||||
|
||||
---
|
||||
|
||||
## 技术可行性分析
|
||||
|
||||
### ✅ Infinite Sessions在BYOK模式下是技术可行的：
|
||||
|
||||
1. **SDK支持**: Copilot SDK允许在任何provider (官方、BYOK、Azure等) 下使用infinite session配置
|
||||
2. **配置独立性**: provider和infinite_sessions配置在SessionConfig中是独立的字段
|
||||
3. **无文档限制**: SDK文档中没有说BYOK模式不支持infinite sessions
|
||||
4. **测试覆盖**: SDK虽然有单独的BYOK测试和infinite-sessions测试，但缺少组合测试
|
||||
|
||||
### ⚠️ 但存在以下设计问题：
|
||||
|
||||
#### 问题1: 意外的自动启用
|
||||
- BYOK模式通常用于**精确控制**自己的API使用
|
||||
- 自动压缩可能会导致**意外的额外请求**和API成本增加
|
||||
- 没有明确的警告或文档说明BYOK也会压缩
|
||||
|
||||
#### 问题2: 没有模式特定的配置
|
||||
```python
|
||||
# 当前实现 - 一刀切
|
||||
if self.valves.INFINITE_SESSION:
|
||||
# 同时应用于官方模型和BYOK模型
|
||||
|
||||
# 应该是 - 模式感知
|
||||
if self.valves.INFINITE_SESSION and not is_byok_model:
|
||||
# 仅对官方模型启用
|
||||
# 或者
|
||||
if self.valves.INFINITE_SESSION_BYOK and is_byok_model:
|
||||
# BYOK专用配置
|
||||
```
|
||||
|
||||
#### 问题3: 压缩质量不确定性
|
||||
- BYOK模型可能是自部署的或开源模型
|
||||
- 上下文压缩由Copilot CLI处理，质量取决于CLI版本
|
||||
- 没有标准化的压缩效果评估
|
||||
|
||||
---
|
||||
|
||||
## 用户报告现象的根本原因
|
||||
|
||||
用户说："BYOK模式本不应该触发压缩，但碰巧用的模型名称与Copilot内置模型相同，结果意外触发了压缩"
|
||||
|
||||
**分析**:
|
||||
1. OpenWebUI Pipe中，infinite_session配置是**全局启用**的 (INFINITE_SESSION=True)
|
||||
2. 模型识别逻辑中，如果模型元数据丢失，会根据模型名称和BYOK活跃状态来推断
|
||||
3. 如果用户使用的BYOK模型名称恰好是 "gpt-4", "claude-3-5-sonnet" 等，可能被识别错误
|
||||
4. 或者用户根本没意识到infinite session在BYOK模式下也被启用了
|
||||
|
||||
---
|
||||
|
||||
## 建议方案
|
||||
|
||||
### 方案1: 保守方案（推荐）
|
||||
**禁用BYOK模式下的automatic compression**
|
||||
|
||||
```python
|
||||
infinite_session_config = None
|
||||
# 只对标准官方模型启用，不对BYOK启用
|
||||
if self.valves.INFINITE_SESSION and not is_byok_model:
|
||||
infinite_session_config = InfiniteSessionConfig(
|
||||
enabled=True,
|
||||
background_compaction_threshold=self.valves.COMPACTION_THRESHOLD,
|
||||
buffer_exhaustion_threshold=self.valves.BUFFER_THRESHOLD,
|
||||
)
|
||||
```
|
||||
|
||||
**优点**:
|
||||
- 尊重BYOK用户的成本控制意愿
|
||||
- 降低意外API使用风险
|
||||
- 与reasoning_effort的BYOK禁用保持一致
|
||||
|
||||
**缺点**: 限制了BYOK用户的功能
|
||||
|
||||
### 方案2: 灵活方案
|
||||
**添加独立的BYOK compression配置**
|
||||
|
||||
```python
|
||||
class Valves(BaseModel):
|
||||
INFINITE_SESSION: bool = Field(
|
||||
default=True,
|
||||
description="Enable Infinite Sessions for standard Copilot models"
|
||||
)
|
||||
INFINITE_SESSION_BYOK: bool = Field(
|
||||
default=False,
|
||||
description="Enable Infinite Sessions for BYOK models (advanced users only)"
|
||||
)
|
||||
|
||||
# 使用逻辑
|
||||
if (self.valves.INFINITE_SESSION and not is_byok_model) or \
|
||||
(self.valves.INFINITE_SESSION_BYOK and is_byok_model):
|
||||
infinite_session_config = InfiniteSessionConfig(...)
|
||||
```
|
||||
|
||||
**优点**:
|
||||
- 给BYOK用户完全控制
|
||||
- 保持向后兼容性
|
||||
- 允许高级用户启用
|
||||
|
||||
**缺点**: 增加配置复杂度
|
||||
|
||||
### 方案3: 警告+ 文档
|
||||
**保持当前实现，但添加文档说明**
|
||||
|
||||
- 在README中明确说明infinite session对所有provider类型都启用
|
||||
- 添加Valve描述提示: "Applies to both standard Copilot and BYOK models"
|
||||
- 在BYOK配置部分明确提到压缩成本
|
||||
|
||||
**优点**: 减少实现负担，给用户知情权
|
||||
|
||||
**缺点**: 对已经启用的用户无帮助
|
||||
|
||||
---
|
||||
|
||||
## 推荐实施
|
||||
|
||||
**优先级**: 高
|
||||
**建议实施方案**: **方案1 (保守方案)** 或 **方案2 (灵活方案)**
|
||||
|
||||
如果选择方案1: 修改line 5063处的条件判断
|
||||
如果选择方案2: 添加INFINITE_SESSION_BYOK配置 + 修改初始化逻辑
|
||||
|
||||
---
|
||||
|
||||
## 相关代码位置
|
||||
|
||||
| 文件 | 行号 | 说明 |
|
||||
|-----|------|------|
|
||||
| `github_copilot_sdk.py` | 364-366 | INFINITE_SESSION Valve定义 |
|
||||
| `github_copilot_sdk.py` | 5063-5069 | Infinite session初始化 |
|
||||
| `github_copilot_sdk.py` | 6199-6220 | is_byok_model判断逻辑 |
|
||||
| `github_copilot_sdk.py` | 6329-6331 | reasoning_effort BYOK处理（参考） |
|
||||
|
||||
---
|
||||
|
||||
## 结论
|
||||
|
||||
**BYOK模式与Infinite Sessions的兼容性**:
|
||||
- ✅ 技术上完全可行
|
||||
- ⚠️ 但存在设计意图不清的问题
|
||||
- ✗ 当前实现对BYOK用户可能不友好
|
||||
|
||||
**推荐**: 实施方案1或2之一，增加BYOK模式的控制粒度。
|
||||
@@ -0,0 +1,295 @@
|
||||
# Client传入和管理分析
|
||||
|
||||
## 当前的Client管理架构
|
||||
|
||||
```
|
||||
┌────────────────────────────────────────┐
|
||||
│ Pipe Instance (github_copilot_sdk.py) │
|
||||
│ │
|
||||
│ _shared_clients = { │
|
||||
│ "token_hash_1": CopilotClient(...), │ ← 基于GitHub Token缓存
|
||||
│ "token_hash_2": CopilotClient(...), │
|
||||
│ } │
|
||||
└────────────────────────────────────────┘
|
||||
│
|
||||
│ await _get_client(token)
|
||||
│
|
||||
▼
|
||||
┌────────────────────────────────────────┐
|
||||
│ CopilotClient Instance │
|
||||
│ │
|
||||
│ [仅需GitHub Token配置] │
|
||||
│ │
|
||||
│ config { │
|
||||
│ github_token: "ghp_...", │
|
||||
│ cli_path: "...", │
|
||||
│ config_dir: "...", │
|
||||
│ env: {...}, │
|
||||
│ cwd: "..." │
|
||||
│ } │
|
||||
└────────────────────────────────────────┘
|
||||
│
|
||||
│ create_session(session_config)
|
||||
│
|
||||
▼
|
||||
┌────────────────────────────────────────┐
|
||||
│ Session (per-session configuration) │
|
||||
│ │
|
||||
│ session_config { │
|
||||
│ model: "real_model_id", │
|
||||
│ provider: { │ ← ⭐ BYOK配置在这里
|
||||
│ type: "openai", │
|
||||
│ base_url: "https://api.openai...",
|
||||
│ api_key: "sk-...", │
|
||||
│ ... │
|
||||
│ }, │
|
||||
│ infinite_sessions: {...}, │
|
||||
│ system_message: {...}, │
|
||||
│ ... │
|
||||
│ } │
|
||||
└────────────────────────────────────────┘
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 目前的流程（代码实际位置）
|
||||
|
||||
### 步骤1：获取或创建Client（line 6208）
|
||||
```python
|
||||
# _pipe_impl中
|
||||
client = await self._get_client(token)
|
||||
```
|
||||
|
||||
### 步骤2：_get_client函数（line 5523-5561）
|
||||
```python
|
||||
async def _get_client(self, token: str) -> Any:
|
||||
"""Get or create the persistent CopilotClient from the pool based on token."""
|
||||
if not token:
|
||||
raise ValueError("GitHub Token is required to initialize CopilotClient")
|
||||
|
||||
token_hash = hashlib.md5(token.encode()).hexdigest()
|
||||
|
||||
# 查看是否已有缓存的client
|
||||
client = self.__class__._shared_clients.get(token_hash)
|
||||
if client and client状态正常:
|
||||
return client # ← 复用已有的client
|
||||
|
||||
# 否则创建新client
|
||||
client_config = self._build_client_config(user_id=None, chat_id=None)
|
||||
client_config["github_token"] = token
|
||||
new_client = CopilotClient(client_config)
|
||||
await new_client.start()
|
||||
self.__class__._shared_clients[token_hash] = new_client
|
||||
return new_client
|
||||
```
|
||||
|
||||
### 步骤3：创建会话时传入provider（line 6253-6270）
|
||||
```python
|
||||
# _pipe_impl中，BYOK部分
|
||||
if is_byok_model:
|
||||
provider_config = {
|
||||
"type": byok_type, # "openai" or "anthropic"
|
||||
"wire_api": byok_wire_api,
|
||||
"base_url": byok_base_url,
|
||||
"api_key": byok_api_key or None,
|
||||
"bearer_token": byok_bearer_token or None,
|
||||
}
|
||||
|
||||
# 然后传入session config
|
||||
session = await client.create_session(config={
|
||||
"model": real_model_id,
|
||||
"provider": provider_config, # ← provider在这里传给session
|
||||
...
|
||||
})
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 关键问题：架构的2个层级
|
||||
|
||||
| 层级 | 用途 | 配置内容 | 缓存方式 |
|
||||
|------|------|---------|---------|
|
||||
| **CopilotClient** | CLI和运行时底层逻辑 | GitHub Token, CLI path, 环境变量 | 基于token_hash全局缓存 |
|
||||
| **Session** | 具体的对话会话 | Model, Provider(BYOK), Tools, System Prompt | 不缓存（每次新建） |
|
||||
|
||||
---
|
||||
|
||||
## 当前的问题
|
||||
|
||||
### 问题1：Client是全局缓存的，但Provider是会话级别的
|
||||
```python
|
||||
# ❓ 如果用户想为不同的BYOK模型使用不同的Client呢？
|
||||
# 当前无法做到，因为Client基于token缓存是全局的
|
||||
|
||||
# 例子：
|
||||
# Client A: OpenAI API key (token_hash_1)
|
||||
# Client B: Anthropic API key (token_hash_2)
|
||||
|
||||
# 但在Pipe中，只有一个GH_TOKEN，导致只能有一个Client
|
||||
```
|
||||
|
||||
### 问题2：Provider和Client是不同的东西
|
||||
```python
|
||||
# CopilotClient = GitHub Copilot SDK客户端
|
||||
# ProviderConfig = OpenAI/Anthropic等的API配置
|
||||
|
||||
# 用户可能混淆：
|
||||
# "怎么传入BYOK的client和provider"
|
||||
# → 实际上只能传provider到session，client是全局的
|
||||
```
|
||||
|
||||
### 问题3：BYOK模型混用的情况处理不清楚
|
||||
```python
|
||||
# 如果用户想在同一个Pipe中：
|
||||
# - Model A 用 OpenAI API
|
||||
# - Model B 用 Anthropic API
|
||||
# - Model C 用自己的本地LLM
|
||||
|
||||
# 当前代码是基于全局BYOK配置的，无法为各模型单独设置
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 改进方案
|
||||
|
||||
### 方案A：保持当前架构，只改Provider映射
|
||||
|
||||
**思路**：Client保持全局（基于GH_TOKEN），但Provider配置基于模型动态选择
|
||||
|
||||
```python
|
||||
# 在Valves中添加
|
||||
class Valves(BaseModel):
|
||||
# ... 现有配置 ...
|
||||
|
||||
# 新增：模型到Provider的映射 (JSON)
|
||||
MODEL_PROVIDER_MAP: str = Field(
|
||||
default="{}",
|
||||
description='Map model IDs to BYOK providers (JSON). Example: '
|
||||
'{"gpt-4": {"type": "openai", "base_url": "...", "api_key": "..."}, '
|
||||
'"claude-3": {"type": "anthropic", "base_url": "...", "api_key": "..."}}'
|
||||
)
|
||||
|
||||
# 在_pipe_impl中
|
||||
def _get_provider_config(self, model_id: str, byok_active: bool) -> Optional[dict]:
|
||||
"""Get provider config for a specific model"""
|
||||
if not byok_active:
|
||||
return None
|
||||
|
||||
try:
|
||||
model_map = json.loads(self.valves.MODEL_PROVIDER_MAP or "{}")
|
||||
return model_map.get(model_id)
|
||||
except:
|
||||
return None
|
||||
|
||||
# 使用时
|
||||
provider_config = self._get_provider_config(real_model_id, byok_active) or {
|
||||
"type": byok_type,
|
||||
"base_url": byok_base_url,
|
||||
"api_key": byok_api_key,
|
||||
...
|
||||
}
|
||||
```
|
||||
|
||||
**优点**：最小改动，复用现有Client架构
|
||||
**缺点**：多个BYOK模型仍共享一个Client（只要GH_TOKEN相同）
|
||||
|
||||
---
|
||||
|
||||
### 方案B：为不同BYOK提供商创建不同的Client
|
||||
|
||||
**思路**：扩展_get_client，支持基于provider_type的多client缓存
|
||||
|
||||
```python
|
||||
async def _get_or_create_client(
|
||||
self,
|
||||
token: str,
|
||||
provider_type: str = "github" # "github", "openai", "anthropic"
|
||||
) -> Any:
|
||||
"""Get or create client based on token and provider type"""
|
||||
|
||||
if provider_type == "github" or not provider_type:
|
||||
# 现有逻辑
|
||||
token_hash = hashlib.md5(token.encode()).hexdigest()
|
||||
else:
|
||||
# 为BYOK提供商创建不同的client
|
||||
composite_key = f"{token}:{provider_type}"
|
||||
token_hash = hashlib.md5(composite_key.encode()).hexdigest()
|
||||
|
||||
# 从缓存获取或创建
|
||||
...
|
||||
```
|
||||
|
||||
**优点**：隔离不同BYOK提供商的Client
|
||||
**缺点**：更复杂，需要更多改动
|
||||
|
||||
---
|
||||
|
||||
## 建议的改进路线
|
||||
|
||||
**优先级1（高）：方案A - 模型到Provider的映射**
|
||||
|
||||
添加Valves配置：
|
||||
```python
|
||||
MODEL_PROVIDER_MAP: str = Field(
|
||||
default="{}",
|
||||
description='Map specific models to their BYOK providers (JSON format)'
|
||||
)
|
||||
```
|
||||
|
||||
使用方式：
|
||||
```
|
||||
{
|
||||
"gpt-4": {
|
||||
"type": "openai",
|
||||
"base_url": "https://api.openai.com/v1",
|
||||
"api_key": "sk-..."
|
||||
},
|
||||
"claude-3": {
|
||||
"type": "anthropic",
|
||||
"base_url": "https://api.anthropic.com/v1",
|
||||
"api_key": "ant-..."
|
||||
},
|
||||
"llama-2": {
|
||||
"type": "openai", # 开源模型通常使用openai兼容API
|
||||
"base_url": "http://localhost:8000/v1",
|
||||
"api_key": "sk-local"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
**优先级2（中）：在_build_session_config中考虑provider_config**
|
||||
|
||||
修改infinite_session初始化，基于provider_config判断：
|
||||
```python
|
||||
def _build_session_config(..., provider_config=None):
|
||||
# 如果使用了BYOK provider，需要特殊处理infinite_session
|
||||
infinite_session_config = None
|
||||
if self.valves.INFINITE_SESSION and provider_config is None:
|
||||
# 仅官方Copilot模型启用compression
|
||||
infinite_session_config = InfiniteSessionConfig(...)
|
||||
```
|
||||
|
||||
**优先级3（低）：方案B - 多client缓存（长期改进）**
|
||||
|
||||
如果需要完全隔离不同BYOK提供商的Client。
|
||||
|
||||
---
|
||||
|
||||
## 总结：如果你要传入BYOK client
|
||||
|
||||
**现状**：
|
||||
- CopilotClient是基于GH_TOKEN全局缓存的
|
||||
- Provider配置是在SessionConfig级别动态设置的
|
||||
- 一个Client可以创建多个Session，每个Session用不同的Provider
|
||||
|
||||
**改进后**：
|
||||
- 添加MODEL_PROVIDER_MAP配置
|
||||
- 对每个模型的请求，动态选择对应的Provider配置
|
||||
- 同一个Client可以为不同Provider服务不同的models
|
||||
|
||||
**你需要做的**：
|
||||
1. 在Valves中配置MODEL_PROVIDER_MAP
|
||||
2. 在模型选择时读取这个映射
|
||||
3. 创建session时用对应的provider_config
|
||||
|
||||
无需修改Client的创建逻辑！
|
||||
@@ -0,0 +1,324 @@
|
||||
# 数据流分析：SDK如何获知用户设计的数据
|
||||
|
||||
## 当前数据流（从OpenWebUI → Pipe → SDK）
|
||||
|
||||
```
|
||||
┌─────────────────────┐
|
||||
│ OpenWebUI UI │
|
||||
│ (用户选择模型) │
|
||||
└──────────┬──────────┘
|
||||
│
|
||||
├─ body.model = "gpt-4"
|
||||
├─ body.messages = [...]
|
||||
├─ __metadata__.base_model_id = ?
|
||||
├─ __metadata__.custom_fields = ?
|
||||
└─ __user__.settings = ?
|
||||
│
|
||||
┌──────────▼──────────┐
|
||||
│ Pipe (github- │
|
||||
│ copilot-sdk.py) │
|
||||
│ │
|
||||
│ 1. 提取model信息 │
|
||||
│ 2. 应用Valves配置 │
|
||||
│ 3. 建立SDK会话 │
|
||||
└──────────┬──────────┘
|
||||
│
|
||||
├─ SessionConfig {
|
||||
│ model: real_model_id
|
||||
│ provider: ProviderConfig (若BYOK)
|
||||
│ infinite_sessions: {...}
|
||||
│ system_message: {...}
|
||||
│ ...
|
||||
│ }
|
||||
│
|
||||
┌──────────▼──────────┐
|
||||
│ Copilot SDK │
|
||||
│ (create_session) │
|
||||
│ │
|
||||
│ 返回:ModelInfo { │
|
||||
│ capabilities { │
|
||||
│ limits { │
|
||||
│ max_context_ │
|
||||
│ window_tokens │
|
||||
│ } │
|
||||
│ } │
|
||||
│ } │
|
||||
└─────────────────────┘
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 关键问题：当前的3个瓶颈
|
||||
|
||||
### 瓶颈1：用户数据的输入点
|
||||
|
||||
**当前支持的输入方式：**
|
||||
|
||||
1. **Valves配置（全局 + 用户级）**
|
||||
```python
|
||||
# 全局设置（Admin）
|
||||
Valves.BYOK_BASE_URL = "https://api.openai.com/v1"
|
||||
Valves.BYOK_API_KEY = "sk-..."
|
||||
|
||||
# 用户级覆盖
|
||||
UserValves.BYOK_API_KEY = "sk-..." (用户自己的key)
|
||||
UserValves.BYOK_BASE_URL = "..."
|
||||
```
|
||||
|
||||
**问题**：无法为特定的BYOK模型设置上下文窗口大小
|
||||
|
||||
2. **__metadata__（来自OpenWebUI）**
|
||||
```python
|
||||
__metadata__ = {
|
||||
"base_model_id": "...",
|
||||
"custom_fields": {...}, # ← 可能包含额外信息
|
||||
"tool_ids": [...],
|
||||
}
|
||||
```
|
||||
|
||||
**问题**：不清楚OpenWebUI是否支持通过metadata传递模型的上下文窗口
|
||||
|
||||
3. **body（来自对话请求）**
|
||||
```python
|
||||
body = {
|
||||
"model": "gpt-4",
|
||||
"messages": [...],
|
||||
"temperature": 0.7,
|
||||
# ← 这里能否添加自定义字段？
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 瓶颈2：模型信息的识别和存储
|
||||
|
||||
**当前代码** (line 5905+)：
|
||||
```python
|
||||
# 解析用户选择的模型
|
||||
request_model = body.get("model", "") # e.g., "gpt-4"
|
||||
real_model_id = request_model
|
||||
|
||||
# 确定实际模型ID
|
||||
base_model_id = _container_get(__metadata__, "base_model_id", "")
|
||||
|
||||
if base_model_id:
|
||||
resolved_id = base_model_id # 使用元数据中的ID
|
||||
else:
|
||||
resolved_id = request_model # 使用用户选择的ID
|
||||
```
|
||||
|
||||
**问题**：
|
||||
- ❌ 没有维护一个"模型元数据缓存"
|
||||
- ❌ 对相同模型的重复请求，每次都需要重新识别
|
||||
- ❌ 不能为特定模型持久化上下文窗口大小
|
||||
|
||||
---
|
||||
|
||||
### 瓶颈3：SDK会话配置的构建
|
||||
|
||||
**当前实现** (line 5058-5100)：
|
||||
```python
|
||||
def _build_session_config(
|
||||
self,
|
||||
real_model_id, # ← 模型ID
|
||||
system_prompt_content,
|
||||
is_streaming=True,
|
||||
is_admin=False,
|
||||
# ... 其他参数
|
||||
):
|
||||
# 无条件地创建infinite session
|
||||
if self.valves.INFINITE_SESSION:
|
||||
infinite_session_config = InfiniteSessionConfig(
|
||||
enabled=True,
|
||||
background_compaction_threshold=self.valves.COMPACTION_THRESHOLD, # 0.80
|
||||
buffer_exhaustion_threshold=self.valves.BUFFER_THRESHOLD, # 0.95
|
||||
)
|
||||
|
||||
# ❌ 这里没有查询该模型的实际上下文窗口大小
|
||||
# ❌ 无法根据模型的真实限制调整压缩阈值
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 解决方案：3个数据流改进步骤
|
||||
|
||||
### 步骤1：添加模型元数据配置（优先级：高）
|
||||
|
||||
在Valves中添加一个**模型元数据映射**：
|
||||
|
||||
```python
|
||||
class Valves(BaseModel):
|
||||
# ... 现有配置 ...
|
||||
|
||||
# 新增：模型上下文窗口映射 (JSON格式)
|
||||
MODEL_CONTEXT_WINDOWS: str = Field(
|
||||
default="{}", # JSON string
|
||||
description='Model context window mapping (JSON). Example: {"gpt-4": 8192, "gpt-4-turbo": 128000, "claude-3": 200000}'
|
||||
)
|
||||
|
||||
# 新增：BYOK模型特定设置 (JSON格式)
|
||||
BYOK_MODEL_CONFIG: str = Field(
|
||||
default="{}", # JSON string
|
||||
description='BYOK-specific model configuration (JSON). Example: {"gpt-4": {"context_window": 8192, "enable_compression": true}}'
|
||||
)
|
||||
```
|
||||
|
||||
**如何使用**：
|
||||
```python
|
||||
# Valves中设置
|
||||
MODEL_CONTEXT_WINDOWS = '{"gpt-4": 8192, "claude-3-5-sonnet": 200000}'
|
||||
|
||||
# Pipe中解析
|
||||
def _get_model_context_window(self, model_id: str) -> Optional[int]:
|
||||
"""从配置中获取模型的上下文窗口大小"""
|
||||
try:
|
||||
config = json.loads(self.valves.MODEL_CONTEXT_WINDOWS or "{}")
|
||||
return config.get(model_id)
|
||||
except:
|
||||
return None
|
||||
```
|
||||
|
||||
### 步骤2：建立模型信息缓存（优先级：中）
|
||||
|
||||
在Pipe中维护一个模型信息缓存：
|
||||
|
||||
```python
|
||||
class Pipe:
|
||||
def __init__(self):
|
||||
# ... 现有代码 ...
|
||||
self._model_info_cache = {} # model_id -> ModelInfo
|
||||
self._context_window_cache = {} # model_id -> context_window_tokens
|
||||
|
||||
def _cache_model_info(self, model_id: str, model_info: ModelInfo):
|
||||
"""缓存SDK返回的模型信息"""
|
||||
self._model_info_cache[model_id] = model_info
|
||||
if model_info.capabilities and model_info.capabilities.limits:
|
||||
self._context_window_cache[model_id] = (
|
||||
model_info.capabilities.limits.max_context_window_tokens
|
||||
)
|
||||
|
||||
def _get_context_window(self, model_id: str) -> Optional[int]:
|
||||
"""获取模型的上下文窗口大小（优先级：SDK > Valves配置 > 默认值）"""
|
||||
# 1. 优先从SDK缓存获取（最可靠）
|
||||
if model_id in self._context_window_cache:
|
||||
return self._context_window_cache[model_id]
|
||||
|
||||
# 2. 其次从Valves配置获取
|
||||
context_window = self._get_model_context_window(model_id)
|
||||
if context_window:
|
||||
return context_window
|
||||
|
||||
# 3. 默认值（未知）
|
||||
return None
|
||||
```
|
||||
|
||||
### 步骤3：使用真实的上下文窗口来优化压缩策略（优先级：中）
|
||||
|
||||
修改_build_session_config：
|
||||
|
||||
```python
|
||||
def _build_session_config(
|
||||
self,
|
||||
real_model_id,
|
||||
# ... 其他参数 ...
|
||||
**kwargs
|
||||
):
|
||||
# 获取模型的真实上下文窗口大小
|
||||
actual_context_window = self._get_context_window(real_model_id)
|
||||
|
||||
# 只对有明确上下文窗口的模型启用压缩
|
||||
infinite_session_config = None
|
||||
if self.valves.INFINITE_SESSION and actual_context_window:
|
||||
# 现在压缩阈值有了明确的含义
|
||||
infinite_session_config = InfiniteSessionConfig(
|
||||
enabled=True,
|
||||
# 80% of actual context window
|
||||
background_compaction_threshold=self.valves.COMPACTION_THRESHOLD,
|
||||
# 95% of actual context window
|
||||
buffer_exhaustion_threshold=self.valves.BUFFER_THRESHOLD,
|
||||
)
|
||||
|
||||
await self._emit_debug_log(
|
||||
f"Infinite Session: model_context={actual_context_window}tokens, "
|
||||
f"compaction_triggers_at={int(actual_context_window * self.valves.COMPACTION_THRESHOLD)}, "
|
||||
f"buffer_triggers_at={int(actual_context_window * self.valves.BUFFER_THRESHOLD)}",
|
||||
__event_call__,
|
||||
)
|
||||
elif self.valves.INFINITE_SESSION and not actual_context_window:
|
||||
logger.warning(
|
||||
f"Infinite Session: Unknown context window for {real_model_id}, "
|
||||
f"compression disabled. Set MODEL_CONTEXT_WINDOWS in Valves to enable."
|
||||
)
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 具体的配置示例
|
||||
|
||||
### 例子1：用户配置BYOK模型的上下文窗口
|
||||
|
||||
**Valves设置**：
|
||||
```
|
||||
MODEL_CONTEXT_WINDOWS = {
|
||||
"gpt-4": 8192,
|
||||
"gpt-4-turbo": 128000,
|
||||
"gpt-4o": 128000,
|
||||
"claude-3": 200000,
|
||||
"claude-3.5-sonnet": 200000,
|
||||
"llama-2-70b": 4096
|
||||
}
|
||||
```
|
||||
|
||||
**效果**：
|
||||
- Pipe会知道"gpt-4"的上下文是8192 tokens
|
||||
- 压缩会在 ~6553 tokens (80%) 时触发
|
||||
- 缓冲会在 ~7782 tokens (95%) 时阻塞
|
||||
|
||||
### 例子2：为特定BYOK模型启用/禁用压缩
|
||||
|
||||
**Valves设置**：
|
||||
```
|
||||
BYOK_MODEL_CONFIG = {
|
||||
"gpt-4": {
|
||||
"context_window": 8192,
|
||||
"enable_infinite_session": true,
|
||||
"compaction_threshold": 0.75
|
||||
},
|
||||
"llama-2-70b": {
|
||||
"context_window": 4096,
|
||||
"enable_infinite_session": false # 禁用压缩
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
**Pipe逻辑**：
|
||||
```python
|
||||
# 检查模型特定的压缩设置
|
||||
def _get_compression_enabled(self, model_id: str) -> bool:
|
||||
try:
|
||||
config = json.loads(self.valves.BYOK_MODEL_CONFIG or "{}")
|
||||
model_config = config.get(model_id, {})
|
||||
return model_config.get("enable_infinite_session", self.valves.INFINITE_SESSION)
|
||||
except:
|
||||
return self.valves.INFINITE_SESSION
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 总结：SDK如何获知用户设计的数据
|
||||
|
||||
| 来源 | 方式 | 更新 | 示例 |
|
||||
|------|------|------|------|
|
||||
| **Valves** | 全局配置 | Admin提前设置 | `MODEL_CONTEXT_WINDOWS` JSON |
|
||||
| **SDK** | SessionConfig返回 | 每次会话创建 | `model_info.capabilities.limits` |
|
||||
| **缓存** | Pipe本地存储 | 首次获取后缓存 | `_context_window_cache` |
|
||||
| **__metadata__** | OpenWebUI传递 | 每次请求随带 | `base_model_id`, custom fields |
|
||||
|
||||
**流程**：
|
||||
1. 用户在Valves中配置 `MODEL_CONTEXT_WINDOWS`
|
||||
2. Pipe在session创建时获取SDK返回的model_info
|
||||
3. Pipe缓存上下文窗口大小
|
||||
4. Pipe根据真实窗口大小调整infinite session的阈值
|
||||
5. SDK使用正确的压缩策略
|
||||
|
||||
这样，**SDK完全知道用户设计的数据**，而无需任何修改SDK本身。
|
||||
@@ -0,0 +1,163 @@
|
||||
# SDK中的上下文限制信息
|
||||
|
||||
## SDK类型定义
|
||||
|
||||
### 1. ModelLimits（copilot-sdk/python/copilot/types.py, line 761-789）
|
||||
|
||||
```python
|
||||
@dataclass
|
||||
class ModelLimits:
|
||||
"""Model limits"""
|
||||
|
||||
max_prompt_tokens: int | None = None # 最大提示符tokens
|
||||
max_context_window_tokens: int | None = None # 最大上下文窗口tokens
|
||||
vision: ModelVisionLimits | None = None # 视觉相关限制
|
||||
```
|
||||
|
||||
### 2. ModelCapabilities（line 817-843）
|
||||
|
||||
```python
|
||||
@dataclass
|
||||
class ModelCapabilities:
|
||||
"""Model capabilities and limits"""
|
||||
|
||||
supports: ModelSupports # 支持的功能（vision, reasoning_effort等）
|
||||
limits: ModelLimits # 上下文和token限制
|
||||
```
|
||||
|
||||
### 3. ModelInfo（line 889-949）
|
||||
|
||||
```python
|
||||
@dataclass
|
||||
class ModelInfo:
|
||||
"""Information about an available model"""
|
||||
|
||||
id: str
|
||||
name: str
|
||||
capabilities: ModelCapabilities # ← 包含limits信息
|
||||
policy: ModelPolicy | None = None
|
||||
billing: ModelBilling | None = None
|
||||
supported_reasoning_efforts: list[str] | None = None
|
||||
default_reasoning_effort: str | None = None
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 关键发现
|
||||
|
||||
### ✅ SDK提供的信息
|
||||
- `model.capabilities.limits.max_context_window_tokens` - 模型的上下文窗口大小
|
||||
- `model.capabilities.limits.max_prompt_tokens` - 最大提示符tokens
|
||||
|
||||
### ❌ OpenWebUI Pipe中的问题
|
||||
**目前Pipe完全没有使用这些信息！**
|
||||
|
||||
在 `github_copilot_sdk.py` 中搜索 `max_context_window`, `capabilities`, `limits` 等，结果为空。
|
||||
|
||||
---
|
||||
|
||||
## 这对BYOK意味着什么？
|
||||
|
||||
### 问题1: BYOK模型的上下文限制未知
|
||||
```python
|
||||
# BYOK模型的capabilities来自哪里？
|
||||
if is_byok_model:
|
||||
# ❓ BYOK模型没有能力信息返回吗？
|
||||
# ❓ 如何知道它的max_context_window_tokens？
|
||||
pass
|
||||
```
|
||||
|
||||
### 问题2: Infinite Session的阈值是硬编码的
|
||||
```python
|
||||
COMPACTION_THRESHOLD: float = Field(
|
||||
default=0.80, # 80%时触发后台压缩
|
||||
description="Background compaction threshold (0.0-1.0)"
|
||||
)
|
||||
BUFFER_THRESHOLD: float = Field(
|
||||
default=0.95, # 95%时阻塞直到压缩完成
|
||||
description="Buffer exhaustion threshold (0.0-1.0)"
|
||||
)
|
||||
|
||||
# 但是 0.80 和 0.95 是什么的百分比？
|
||||
# - 是模型的max_context_window_tokens吗？
|
||||
# - 还是固定的某个值？
|
||||
# - BYOK模型的上下文窗口可能完全不同！
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 改进方向
|
||||
|
||||
### 方案A: 利用SDK提供的模型限制信息
|
||||
```python
|
||||
# 在获取模型信息时，保存capabilities
|
||||
self._model_capabilities = model_info.capabilities
|
||||
|
||||
# 在初始化infinite session时，使用实际的上下文窗口
|
||||
if model_info.capabilities.limits.max_context_window_tokens:
|
||||
actual_context_window = model_info.capabilities.limits.max_context_window_tokens
|
||||
|
||||
# 动态调整压缩阈值而不是固定值
|
||||
compaction_threshold = self.valves.COMPACTION_THRESHOLD
|
||||
buffer_threshold = self.valves.BUFFER_THRESHOLD
|
||||
# 这些现在有了明确的含义：是模型实际上下文窗口大小的百分比
|
||||
```
|
||||
|
||||
### 方案B: BYOK模型的显式配置
|
||||
如果BYOK模型不提供capabilities信息，需要用户手动设置：
|
||||
|
||||
```python
|
||||
class Valves(BaseModel):
|
||||
# ... existing config ...
|
||||
|
||||
BYOK_CONTEXT_WINDOW: int = Field(
|
||||
default=0, # 0表示自动检测或禁用compression
|
||||
description="Manual context window size for BYOK models (tokens). 0=auto-detect or disabled"
|
||||
)
|
||||
|
||||
BYOK_INFINITE_SESSION: bool = Field(
|
||||
default=False,
|
||||
description="Enable infinite sessions for BYOK models (requires BYOK_CONTEXT_WINDOW > 0)"
|
||||
)
|
||||
```
|
||||
|
||||
### 方案C: 从会话反馈中学习（最可靠）
|
||||
```python
|
||||
# infinite session压缩完成时，获取实际的context window使用情况
|
||||
# (需要SDK或CLI提供反馈)
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 建议实施路线
|
||||
|
||||
**优先级1（必须）**: 检查BYOK模式下是否能获取capabilities
|
||||
```python
|
||||
# 测试代码
|
||||
if is_byok_model:
|
||||
# 发送一个测试请求，看是否能从响应中获取model capabilities
|
||||
session = await client.create_session(config=session_config)
|
||||
# session是否包含model info？
|
||||
# 能否访问session.model_capabilities？
|
||||
```
|
||||
|
||||
**优先级2（重要）**: 如果BYOK没有capabilities，添加手动配置
|
||||
```python
|
||||
# 在BYOK配置中添加context_window字段
|
||||
BYOK_CONTEXT_WINDOW: int = Field(default=0)
|
||||
```
|
||||
|
||||
**优先级3（长期）**: 利用真实的上下文窗口来调整压缩策略
|
||||
```python
|
||||
# 而不是单纯的百分比，使用实际的token数
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 关键问题列表
|
||||
|
||||
1. [ ] BYOK模型在create_session后能否获取capabilities信息？
|
||||
2. [ ] 如果能获取，max_context_window_tokens的值是否准确？
|
||||
3. [ ] 如果不能获取，是否需要用户手动提供？
|
||||
4. [ ] 当前的0.80/0.95阈值是否对所有模型都适用？
|
||||
5. [ ] 不同的BYOK提供商(OpenAI vs Anthropic)的上下文窗口差异有多大？
|
||||
Reference in New Issue
Block a user