mirror of
https://github.com/csunny/DB-GPT.git
synced 2026-07-16 17:15:22 +00:00
doc:add react agent context manager,including compact, budget.
This commit is contained in:
aries_ckt
317
docs/docs/agents/modules/context-management.md
Normal file
317
docs/docs/agents/modules/context-management.md
Normal file
@@ -0,0 +1,317 @@
|
||||
---
|
||||
title: Context Management
|
||||
---
|
||||
|
||||
# Agent Context Management
|
||||
|
||||
Agent context management keeps long-running ReAct conversations inside the
|
||||
model context window without losing the working state of the task. It tracks
|
||||
token usage before each model call, emits live context status events, and applies
|
||||
progressively stronger compaction when the conversation grows too large.
|
||||
|
||||
## Overview
|
||||
|
||||
```text
|
||||
User task
|
||||
|
|
||||
v
|
||||
Agent builds messages
|
||||
system prompt + task progress + memory + recent ReAct rounds
|
||||
|
|
||||
v
|
||||
Count tokens
|
||||
ProxyTokenizerWrapper.count_token(model_name)
|
||||
fallback: len(content) // 4
|
||||
|
|
||||
v
|
||||
Compute budget
|
||||
effective_budget = max_context_tokens - reserved_tokens
|
||||
usage_ratio = used_tokens / effective_budget
|
||||
|
|
||||
v
|
||||
Classify state
|
||||
normal < warning < error < critical < overflow
|
||||
|
|
||||
+-- normal --------------------------------------+
|
||||
| |
|
||||
v |
|
||||
Send messages to LLM |
|
||||
|
|
||||
warning or above |
|
||||
| |
|
||||
v |
|
||||
Layer 1: Observation micro-compaction |
|
||||
truncate old tool observations |
|
||||
| |
|
||||
v |
|
||||
Recount and emit context.status |
|
||||
| |
|
||||
+-- below warning -------------------------------+
|
||||
| |
|
||||
v |
|
||||
Layer 2: Session memory compaction |
|
||||
drop old ReAct rounds, keep recent rounds |
|
||||
| |
|
||||
v |
|
||||
Recount and emit context.status |
|
||||
| |
|
||||
+-- below error ---------------------------------+
|
||||
| |
|
||||
v |
|
||||
Layer 3: Full context compression |
|
||||
summarize old rounds with the LLM |
|
||||
| |
|
||||
v |
|
||||
Recount and emit context.status |
|
||||
| |
|
||||
+----------------------------------------------->+
|
||||
|
||||
If the LLM still returns a context overflow error:
|
||||
|
||||
LLM context_too_long / maximum context length error
|
||||
|
|
||||
v
|
||||
Layer 4: Reactive compaction
|
||||
keep system prompt + last 2 ReAct rounds
|
||||
|
|
||||
v
|
||||
Retry the model call once with the compacted messages
|
||||
```
|
||||
|
||||
Tool results are preserved through a separate snapshot path:
|
||||
|
||||
```text
|
||||
Action succeeds
|
||||
|
|
||||
v
|
||||
Write full operation snapshot
|
||||
step, action, action_input, observation, thought, timestamp
|
||||
|
|
||||
v
|
||||
Store snapshot path on the memory fragment
|
||||
and in task progress metadata
|
||||
|
|
||||
v
|
||||
Rebuild memory for future prompts
|
||||
Observation: short or compacted observation
|
||||
[Full detail available at: /path/to/snapshot.json]
|
||||
|
|
||||
v
|
||||
Layer 1 / Layer 2 can shrink prompt text
|
||||
without deleting the original tool result file
|
||||
```
|
||||
|
||||
## Token Budget
|
||||
|
||||
The context manager counts the tokens in the current `AgentMessage` list before
|
||||
the model call. Counting uses `ProxyTokenizerWrapper` with the active
|
||||
`model_name`. If the tokenizer cannot count the content, DB-GPT falls back to a
|
||||
rough estimate of four characters per token.
|
||||
|
||||
The usable context window is:
|
||||
|
||||
```text
|
||||
effective_budget = max_context_tokens - reserved_tokens
|
||||
```
|
||||
|
||||
`reserved_tokens` keeps space for the model response so the prompt does not fill
|
||||
the entire model window.
|
||||
|
||||
## States And Thresholds
|
||||
|
||||
| State | Default trigger | Meaning |
|
||||
| --- | --- | --- |
|
||||
| `normal` | `< 70%` | No compaction. |
|
||||
| `warning` | `>= 70%` | Start lightweight compaction. |
|
||||
| `error` | `>= 90%` | Use LLM-based summary compaction when needed. |
|
||||
| `critical` | `>= 95%` | Same as error, but reported as a more urgent state. |
|
||||
| `overflow` | `>= 100%` | Prompt is over the effective budget. |
|
||||
|
||||
After every count and every compaction layer, the backend emits a
|
||||
`context.status` event with:
|
||||
|
||||
```json
|
||||
{
|
||||
"type": "context.status",
|
||||
"used": 19000,
|
||||
"budget": 115904,
|
||||
"ratio": 0.164,
|
||||
"state": "normal",
|
||||
"compact_layer": null
|
||||
}
|
||||
```
|
||||
|
||||
The UI renders this as a compact context-window indicator.
|
||||
|
||||
## Compaction Layers
|
||||
|
||||
### Layer 1: Observation Micro-Compaction
|
||||
|
||||
Layer 1 is the lightest compaction. It only shortens old `Observation:` messages
|
||||
from tool calls. Recent rounds are preserved in full.
|
||||
|
||||
Rules:
|
||||
|
||||
- Triggered when usage reaches `warning_threshold`.
|
||||
- A round is considered old when it is older than
|
||||
`max_observation_age_rounds`.
|
||||
- Old observations are truncated to `truncated_observation_max_chars`.
|
||||
- If the observation has a snapshot path, the compacted message keeps a pointer
|
||||
to the full detail.
|
||||
|
||||
This layer is cheap and deterministic. It does not call the LLM.
|
||||
|
||||
### Layer 2: Session Memory Compaction
|
||||
|
||||
Layer 2 removes old complete ReAct rounds from the prompt. It relies on the
|
||||
task-progress summary already injected into the system prompt, so the agent still
|
||||
knows what has been completed.
|
||||
|
||||
Rules:
|
||||
|
||||
- Triggered when the prompt is still at or above `warning_threshold` after Layer
|
||||
1.
|
||||
- Always keeps at least `min_keep_recent_rounds`.
|
||||
- Also keeps enough recent content to satisfy `min_keep_tokens`.
|
||||
- Drops complete old rounds rather than arbitrary individual messages.
|
||||
|
||||
This layer is also deterministic and does not call the LLM.
|
||||
|
||||
### Layer 3: Full Context Compression
|
||||
|
||||
Layer 3 summarizes old conversation rounds into a structured context summary
|
||||
with the LLM, then keeps that summary plus the recent rounds.
|
||||
|
||||
Rules:
|
||||
|
||||
- Triggered when usage is at or above `error_threshold`.
|
||||
- Keeps the last `min_keep_recent_rounds` unchanged.
|
||||
- Summarizes older messages into one synthetic summary message.
|
||||
- The summary prompt asks the model to preserve exact task state, paths, values,
|
||||
variable names, errors, and next steps.
|
||||
- If summarization fails repeatedly, a circuit breaker stops retrying after
|
||||
`max_compact_failures`.
|
||||
|
||||
This layer is more expensive, but it preserves more semantic continuity than
|
||||
simply dropping old messages.
|
||||
|
||||
### Layer 4: Reactive Compaction
|
||||
|
||||
Layer 4 is an emergency path. It is not triggered by the normal budget state
|
||||
machine. Instead, it runs when the model call fails with a context overflow
|
||||
error such as `context_too_long`, `context_length_exceeded`, or
|
||||
`maximum context length`.
|
||||
|
||||
Rules:
|
||||
|
||||
- Keeps system messages.
|
||||
- Keeps only the last two ReAct rounds.
|
||||
- Relies on the task-progress summary in the system prompt to preserve task
|
||||
continuity.
|
||||
- Retries the model call once with the compacted messages.
|
||||
|
||||
This layer is intentionally aggressive because it is only used after the model
|
||||
has already rejected the prompt.
|
||||
|
||||
## Tool Result Snapshots
|
||||
|
||||
Tool observations can be large: SQL result tables, generated code output,
|
||||
interpreter logs, file paths, report metadata, and intermediate computed values
|
||||
may quickly dominate the prompt. DB-GPT keeps the prompt compact by separating
|
||||
the full operation detail from the text that must stay in the model context.
|
||||
|
||||
When an action succeeds, the agent writes a JSON snapshot for the full operation.
|
||||
The snapshot includes:
|
||||
|
||||
- `step`
|
||||
- `action`
|
||||
- `phase`
|
||||
- `action_intention`
|
||||
- `action_reason`
|
||||
- `thought`
|
||||
- `action_input`
|
||||
- `observation`
|
||||
- `timestamp`
|
||||
- `conv_id`
|
||||
|
||||
By default, snapshots are written under:
|
||||
|
||||
```text
|
||||
$DBGPT_HOME/workspace/op_snapshots/<conv_id>/
|
||||
```
|
||||
|
||||
If `AgentContext.output_dir` is set, DB-GPT uses that directory instead.
|
||||
|
||||
Each snapshot file is named by step and action:
|
||||
|
||||
```text
|
||||
step_003_sql_query.json
|
||||
step_006_code_interpreter.json
|
||||
```
|
||||
|
||||
The snapshot path is attached to the in-memory `AgentMemoryFragment` and also
|
||||
recorded in the task-progress metadata. When the agent later rebuilds memories
|
||||
into prompt messages, it appends a lightweight reference:
|
||||
|
||||
```text
|
||||
Observation: <observation text>
|
||||
[Full detail available at: /path/to/step_003_sql_query.json]
|
||||
```
|
||||
|
||||
This matters during compaction:
|
||||
|
||||
- Layer 1 may truncate old `Observation:` text, but it preserves the snapshot
|
||||
reference when available.
|
||||
- Layer 2 may remove old ReAct rounds from the prompt, but task progress still
|
||||
records the snapshot filename as a reference.
|
||||
- Layer 3 summarizes old messages, while the original tool result remains on
|
||||
disk for exact recovery.
|
||||
|
||||
In other words, compaction reduces the prompt payload; it does not have to be
|
||||
the only place where exact tool output lives.
|
||||
|
||||
## Configuration
|
||||
|
||||
Agent context management can be configured in the application TOML file:
|
||||
|
||||
```toml
|
||||
[service.web.agent_context]
|
||||
# Set to 0 to auto-detect from the selected model metadata.
|
||||
max_context_tokens = 0
|
||||
reserved_tokens = 4096
|
||||
warning_threshold = 0.70
|
||||
error_threshold = 0.90
|
||||
critical_threshold = 0.95
|
||||
min_keep_recent_rounds = 3
|
||||
max_observation_age_rounds = 5
|
||||
truncated_observation_max_chars = 200
|
||||
min_keep_tokens = 10000
|
||||
max_compact_failures = 3
|
||||
```
|
||||
|
||||
When `max_context_tokens` is `0`, DB-GPT tries to read the selected model's
|
||||
`context_length` from `llm_client.get_model_metadata(model_name)`. If the model
|
||||
metadata is unavailable, it falls back to the default budget.
|
||||
|
||||
For stable behavior, set `context_length` on each LLM deployment:
|
||||
|
||||
```toml
|
||||
[[models.llms]]
|
||||
name = "Qwen/Qwen2.5-Coder-32B-Instruct"
|
||||
provider = "proxy/siliconflow"
|
||||
api_key = "${env:SILICONFLOW_API_KEY}"
|
||||
context_length = 32768
|
||||
```
|
||||
|
||||
With this setup, switching models also switches the effective context budget.
|
||||
|
||||
## Design Notes
|
||||
|
||||
- Layer 1 and Layer 2 are deterministic and cheap. They are preferred before
|
||||
any LLM summarization.
|
||||
- Layer 3 uses the LLM only when the context is close to failure.
|
||||
- Layer 4 is a last-resort retry path for model-side context overflow errors.
|
||||
- The frontend receives `context.status` events independently from normal chat
|
||||
text, so UI indicators can update without polluting the conversation.
|
||||
- Compaction is progressive: after each layer, DB-GPT recounts tokens and stops
|
||||
escalating if the prompt returns to a safe state.
|
||||
@@ -450,5 +450,9 @@
|
||||
"sidebar.docsSidebar.doc.Use Custom Skills": {
|
||||
"message": "使用自定义技能",
|
||||
"description": "The label for the doc item Use Custom Skills in sidebar docsSidebar, linking to the doc dbgpts/how-to-use-custom-skill"
|
||||
},
|
||||
"sidebar.docsSidebar.doc.Context Management": {
|
||||
"message": "上下文管理",
|
||||
"description": "The label for the doc item Context Management in sidebar docsSidebar, linking to the doc agents/modules/context-management"
|
||||
}
|
||||
}
|
||||
|
||||
@@ -0,0 +1,279 @@
|
||||
---
|
||||
title: 上下文管理
|
||||
---
|
||||
|
||||
# Agent 上下文管理
|
||||
|
||||
Agent 上下文管理用于让长时间运行的 ReAct 对话稳定保持在模型上下文窗口内，同时尽量不丢失任务状态。它会在每次调用模型前统计 token 使用量，向前端发送实时上下文状态，并在对话变长时按层级逐步压缩。
|
||||
|
||||
## 总览
|
||||
|
||||
```text
|
||||
用户任务
|
||||
|
|
||||
v
|
||||
Agent 构造消息
|
||||
system prompt + task progress + memory + 最近 ReAct 轮次
|
||||
|
|
||||
v
|
||||
统计 token
|
||||
ProxyTokenizerWrapper.count_token(model_name)
|
||||
兜底估算：len(content) // 4
|
||||
|
|
||||
v
|
||||
计算预算
|
||||
effective_budget = max_context_tokens - reserved_tokens
|
||||
usage_ratio = used_tokens / effective_budget
|
||||
|
|
||||
v
|
||||
判断状态
|
||||
normal < warning < error < critical < overflow
|
||||
|
|
||||
+-- normal --------------------------------------+
|
||||
| |
|
||||
v |
|
||||
发送消息给 LLM |
|
||||
|
|
||||
warning 及以上 |
|
||||
| |
|
||||
v |
|
||||
第 1 层：Observation 微压缩 |
|
||||
截断较旧的工具观察结果 |
|
||||
| |
|
||||
v |
|
||||
重新统计并发送 context.status |
|
||||
| |
|
||||
+-- 低于 warning -------------------------------+
|
||||
| |
|
||||
v |
|
||||
第 2 层：Session memory 压缩 |
|
||||
丢弃较旧 ReAct 轮次，保留最近轮次 |
|
||||
| |
|
||||
v |
|
||||
重新统计并发送 context.status |
|
||||
| |
|
||||
+-- 低于 error ---------------------------------+
|
||||
| |
|
||||
v |
|
||||
第 3 层：完整上下文压缩 |
|
||||
用 LLM 总结旧轮次 |
|
||||
| |
|
||||
v |
|
||||
重新统计并发送 context.status |
|
||||
| |
|
||||
+----------------------------------------------->+
|
||||
|
||||
如果模型仍然返回上下文溢出错误：
|
||||
|
||||
LLM context_too_long / maximum context length error
|
||||
|
|
||||
v
|
||||
第 4 层：Reactive 压缩
|
||||
保留 system prompt + 最后 2 个 ReAct 轮次
|
||||
|
|
||||
v
|
||||
用压缩后的消息重试一次模型调用
|
||||
```
|
||||
|
||||
工具结果会通过单独的快照路径保留：
|
||||
|
||||
```text
|
||||
Action 执行成功
|
||||
|
|
||||
v
|
||||
写入完整操作快照
|
||||
step、action、action_input、observation、thought、timestamp
|
||||
|
|
||||
v
|
||||
把 snapshot path 存到 memory fragment
|
||||
以及 task progress metadata
|
||||
|
|
||||
v
|
||||
未来重建 memory prompt
|
||||
Observation: 短观察结果或已压缩观察结果
|
||||
[Full detail available at: /path/to/snapshot.json]
|
||||
|
|
||||
v
|
||||
第 1 层 / 第 2 层可以缩小 prompt 文本
|
||||
但不会删除原始工具结果文件
|
||||
```
|
||||
|
||||
## Token 预算
|
||||
|
||||
上下文管理器会在模型调用前统计当前 `AgentMessage` 列表的 token 数。统计逻辑使用 `ProxyTokenizerWrapper`，并传入当前 `model_name`。如果 tokenizer 无法统计，则退化为粗略估算：每 4 个字符按 1 个 token 计算。
|
||||
|
||||
可用上下文窗口计算方式：
|
||||
|
||||
```text
|
||||
effective_budget = max_context_tokens - reserved_tokens
|
||||
```
|
||||
|
||||
`reserved_tokens` 用于为模型输出预留空间，避免 prompt 本身占满整个模型窗口。
|
||||
|
||||
## 状态与阈值
|
||||
|
||||
| 状态 | 默认触发条件 | 含义 |
|
||||
| --- | --- | --- |
|
||||
| `normal` | `< 70%` | 不压缩。 |
|
||||
| `warning` | `>= 70%` | 开始轻量压缩。 |
|
||||
| `error` | `>= 90%` | 必要时使用 LLM 总结压缩。 |
|
||||
| `critical` | `>= 95%` | 与 error 类似，但表示更紧急。 |
|
||||
| `overflow` | `>= 100%` | prompt 超过有效预算。 |
|
||||
|
||||
每次统计和每层压缩后，后端都会发送 `context.status` 事件：
|
||||
|
||||
```json
|
||||
{
|
||||
"type": "context.status",
|
||||
"used": 19000,
|
||||
"budget": 115904,
|
||||
"ratio": 0.164,
|
||||
"state": "normal",
|
||||
"compact_layer": null
|
||||
}
|
||||
```
|
||||
|
||||
前端会把这个事件展示为上下文窗口指示器。
|
||||
|
||||
## 压缩层级
|
||||
|
||||
### 第 1 层：Observation 微压缩
|
||||
|
||||
第 1 层是最轻量的压缩，只处理旧工具调用产生的 `Observation:` 消息。最近轮次会完整保留。
|
||||
|
||||
规则：
|
||||
|
||||
- 当使用率达到 `warning_threshold` 时触发。
|
||||
- 超过 `max_observation_age_rounds` 的轮次会被视为旧轮次。
|
||||
- 旧 Observation 会被截断到 `truncated_observation_max_chars`。
|
||||
- 如果 Observation 有快照路径，压缩后的消息会保留完整详情的引用。
|
||||
|
||||
这一层成本低、可确定，不调用 LLM。
|
||||
|
||||
### 第 2 层：Session Memory 压缩
|
||||
|
||||
第 2 层会从 prompt 中移除旧的完整 ReAct 轮次。它依赖已经注入 system prompt 的 task-progress summary，因此 Agent 仍然知道哪些工作已经完成。
|
||||
|
||||
规则：
|
||||
|
||||
- 当第 1 层后仍然达到或超过 `warning_threshold` 时触发。
|
||||
- 至少保留 `min_keep_recent_rounds` 个最近轮次。
|
||||
- 同时尽量保留不少于 `min_keep_tokens` 的最近内容。
|
||||
- 丢弃的是完整旧轮次，而不是任意切分单条消息。
|
||||
|
||||
这一层也是确定性逻辑，不调用 LLM。
|
||||
|
||||
### 第 3 层：完整上下文压缩
|
||||
|
||||
第 3 层会用 LLM 把较旧的对话轮次压缩成结构化上下文摘要，然后保留摘要和最近轮次。
|
||||
|
||||
规则：
|
||||
|
||||
- 当使用率达到或超过 `error_threshold` 时触发。
|
||||
- 最近 `min_keep_recent_rounds` 个轮次原样保留。
|
||||
- 更旧的消息会被总结成一条合成摘要消息。
|
||||
- 摘要提示词会要求模型保留精确的任务状态、路径、数值、变量名、错误和下一步。
|
||||
- 如果总结连续失败，达到 `max_compact_failures` 后会触发熔断，停止继续尝试。
|
||||
|
||||
这一层成本更高，但比直接丢弃旧消息更能保持语义连续性。
|
||||
|
||||
### 第 4 层：Reactive 压缩
|
||||
|
||||
第 4 层是紧急兜底路径。它不是由正常预算状态机触发，而是在模型调用返回上下文溢出错误时触发，例如 `context_too_long`、`context_length_exceeded` 或 `maximum context length`。
|
||||
|
||||
规则：
|
||||
|
||||
- 保留 system 消息。
|
||||
- 只保留最后 2 个 ReAct 轮次。
|
||||
- 依赖 system prompt 中的 task-progress summary 保持任务连续性。
|
||||
- 用压缩后的消息重试一次模型调用。
|
||||
|
||||
这一层非常激进，因为它只在模型已经拒绝当前 prompt 后才使用。
|
||||
|
||||
## 工具结果快照
|
||||
|
||||
工具观察结果可能非常大：SQL 结果表、生成代码输出、解释器日志、文件路径、报告元数据和中间计算值都可能快速占满 prompt。DB-GPT 通过把完整操作详情和需要进入模型上下文的文本拆开，来保持 prompt 紧凑。
|
||||
|
||||
当一个 action 成功执行后，Agent 会为完整操作写入一份 JSON 快照。快照包含：
|
||||
|
||||
- `step`
|
||||
- `action`
|
||||
- `phase`
|
||||
- `action_intention`
|
||||
- `action_reason`
|
||||
- `thought`
|
||||
- `action_input`
|
||||
- `observation`
|
||||
- `timestamp`
|
||||
- `conv_id`
|
||||
|
||||
默认快照目录为：
|
||||
|
||||
```text
|
||||
$DBGPT_HOME/workspace/op_snapshots/<conv_id>/
|
||||
```
|
||||
|
||||
如果设置了 `AgentContext.output_dir`，DB-GPT 会优先使用该目录。
|
||||
|
||||
快照文件名由步骤和 action 组成：
|
||||
|
||||
```text
|
||||
step_003_sql_query.json
|
||||
step_006_code_interpreter.json
|
||||
```
|
||||
|
||||
快照路径会挂到内存中的 `AgentMemoryFragment` 上，也会记录到 task-progress metadata 中。当 Agent 后续把 memory 重建为 prompt 消息时，会附加一条轻量引用：
|
||||
|
||||
```text
|
||||
Observation: <observation text>
|
||||
[Full detail available at: /path/to/step_003_sql_query.json]
|
||||
```
|
||||
|
||||
这对压缩很重要：
|
||||
|
||||
- 第 1 层可能会截断旧的 `Observation:` 文本，但会尽量保留快照引用。
|
||||
- 第 2 层可能会把旧 ReAct 轮次从 prompt 中移除，但 task progress 仍会记录快照文件名作为引用。
|
||||
- 第 3 层会总结旧消息，而原始工具结果仍然保留在磁盘上，便于精确恢复。
|
||||
|
||||
换句话说，压缩减少的是 prompt 负载；精确工具结果不必只存在于上下文文本里。
|
||||
|
||||
## 配置
|
||||
|
||||
可以在应用 TOML 文件中配置 Agent 上下文管理：
|
||||
|
||||
```toml
|
||||
[service.web.agent_context]
|
||||
# 设置为 0 时，会从当前模型 metadata 自动读取上下文窗口。
|
||||
max_context_tokens = 0
|
||||
reserved_tokens = 4096
|
||||
warning_threshold = 0.70
|
||||
error_threshold = 0.90
|
||||
critical_threshold = 0.95
|
||||
min_keep_recent_rounds = 3
|
||||
max_observation_age_rounds = 5
|
||||
truncated_observation_max_chars = 200
|
||||
min_keep_tokens = 10000
|
||||
max_compact_failures = 3
|
||||
```
|
||||
|
||||
当 `max_context_tokens` 为 `0` 时，DB-GPT 会尝试通过 `llm_client.get_model_metadata(model_name)` 读取所选模型的 `context_length`。如果模型 metadata 不可用，则回退到默认预算。
|
||||
|
||||
为了让行为稳定，建议在每个 LLM 部署配置中显式设置 `context_length`：
|
||||
|
||||
```toml
|
||||
[[models.llms]]
|
||||
name = "Qwen/Qwen2.5-Coder-32B-Instruct"
|
||||
provider = "proxy/siliconflow"
|
||||
api_key = "${env:SILICONFLOW_API_KEY}"
|
||||
context_length = 32768
|
||||
```
|
||||
|
||||
这样切换模型时，上下文预算也会跟随模型能力自动变化。
|
||||
|
||||
## 设计说明
|
||||
|
||||
- 第 1 层和第 2 层都是确定性、低成本压缩，优先于 LLM 总结。
|
||||
- 第 3 层只在上下文接近失败时调用 LLM。
|
||||
- 第 4 层是模型侧上下文溢出错误后的最后重试路径。
|
||||
- 前端独立接收 `context.status` 事件，因此上下文窗口指示器可以实时更新，不会污染正常对话内容。
|
||||
- 压缩是渐进式的：每一层后都会重新统计 token，如果 prompt 已回到安全状态，就不会继续升级到更强压缩。
|
||||
@@ -218,6 +218,7 @@ const sidebars = {
|
||||
label: "Memory",
|
||||
items: [
|
||||
{ type: "doc", id: "agents/modules/memory/memory" },
|
||||
{ type: "doc", id: "agents/modules/context-management", label: "Context Management" },
|
||||
{ type: "doc", id: "agents/modules/memory/sensory_memory" },
|
||||
{ type: "doc", id: "agents/modules/memory/short_term_memory" },
|
||||
{ type: "doc", id: "agents/modules/memory/long_term_memory" },
|
||||
|
||||
@@ -1374,11 +1374,6 @@ class ConversableAgent(Role, Agent):
|
||||
|
||||
# Multi-layer context management: compress if budget exceeded
|
||||
ctx_mgr: Optional[ContextManager] = getattr(self, "_context_manager", None)
|
||||
logger.warning(
|
||||
"[CTX-DEBUG] _context_manager lookup: agent=%s, found=%s",
|
||||
self.name,
|
||||
ctx_mgr is not None,
|
||||
)
|
||||
if ctx_mgr is not None:
|
||||
agent_messages = await ctx_mgr.manage_context(
|
||||
messages=agent_messages,
|
||||
|
||||
@@ -237,7 +237,9 @@ class FullContextCompression:
|
||||
old_indices.add(i)
|
||||
|
||||
old_msgs = [conv_msgs[i] for i in sorted(old_indices)]
|
||||
recent_msgs = [conv_msgs[i] for i in range(len(conv_msgs)) if i not in old_indices]
|
||||
recent_msgs = [
|
||||
conv_msgs[i] for i in range(len(conv_msgs)) if i not in old_indices
|
||||
]
|
||||
|
||||
# Build conversation text for summarization
|
||||
conv_lines = []
|
||||
|
||||
@@ -109,7 +109,6 @@ const useChat = ({ queryAgentURL = '/api/v1/chat/completions', app_code }: Props
|
||||
let message = event.data;
|
||||
let needReplaceNewline = false;
|
||||
let parsedData;
|
||||
console.log('[CTX-DEBUG] SSE onmessage raw:', message?.substring(0, 200));
|
||||
|
||||
try {
|
||||
parsedData = JSON.parse(message);
|
||||
@@ -119,7 +118,6 @@ const useChat = ({ queryAgentURL = '/api/v1/chat/completions', app_code }: Props
|
||||
// React-agent format: {"type": "context.status", "used": ..., "budget": ..., ...}
|
||||
const cs = parsedData.context_status ?? (parsedData.type === 'context.status' ? parsedData : null);
|
||||
if (cs) {
|
||||
console.log('[CTX-DEBUG] Received context_status SSE event:', cs);
|
||||
// Only show banner when Layer 3 (LLM compression) is active
|
||||
if (cs.compact_layer === 'layer3') {
|
||||
setContextStatus({
|
||||
|
||||
Reference in New Issue
Block a user