fix: 降低 LLM 上下文调用成本风险

astrbot/builtin_stars/astrbot/long_term_memory.py

@@ -9,6 +9,7 @@
 from astrbot.api.message_components import At, Image, Plain
 from astrbot.api.platform import MessageType
 from astrbot.api.provider import LLMResponse, Provider, ProviderRequest
+from astrbot.core.agent.message import TextPart
 from astrbot.core.astrbot_config_mgr import AstrBotConfigManager
 
 """
@@ -17,6 +18,9 @@
 
 
 class LongTermMemory:
+    DEFAULT_MAX_GROUP_MESSAGES = 50
+    DEFAULT_GROUP_ICL_TOKEN_BUDGET = 4000
+
     def __init__(self, acm: AstrBotConfigManager, context: star.Context) -> None:
         self.acm = acm
         self.context = context
@@ -29,7 +33,19 @@ def cfg(self, event: AstrMessageEvent):
             max_cnt = int(cfg["provider_ltm_settings"]["group_message_max_cnt"])
         except BaseException as e:
             logger.error(e)
-            max_cnt = 300
+            max_cnt = self.DEFAULT_MAX_GROUP_MESSAGES
+        max_cnt = max(1, max_cnt)
+        try:
+            group_icl_token_budget = int(
+                cfg["provider_ltm_settings"].get(
+                    "group_icl_token_budget",
+                    self.DEFAULT_GROUP_ICL_TOKEN_BUDGET,
+                )
+            )
+        except BaseException as e:
+            logger.error(e)
+            group_icl_token_budget = self.DEFAULT_GROUP_ICL_TOKEN_BUDGET
+        group_icl_token_budget = max(1, group_icl_token_budget)
         image_caption_prompt = cfg["provider_settings"]["image_caption_prompt"]
         image_caption_provider_id = cfg["provider_ltm_settings"].get(
             "image_caption_provider_id"
@@ -45,6 +61,7 @@ def cfg(self, event: AstrMessageEvent):
         ar_whitelist = active_reply.get("whitelist", [])
         ret = {
             "max_cnt": max_cnt,
+            "group_icl_token_budget": group_icl_token_budget,
             "image_caption": image_caption,
             "image_caption_prompt": image_caption_prompt,
             "image_caption_provider_id": image_caption_provider_id,
@@ -56,6 +73,74 @@ def cfg(self, event: AstrMessageEvent):
         }
         return ret
 
+    @staticmethod
+    def _estimate_text_tokens(text: str) -> int:
+        chinese_count = len([c for c in text if "\u4e00" <= c <= "\u9fff"])
+        other_count = len(text) - chinese_count
+        return int(chinese_count * 0.6 + other_count * 0.3)
+
+    def _trim_text_to_token_budget(self, text: str, token_budget: int) -> str:
+        marker = "[truncated]\n"
+        marker_tokens = self._estimate_text_tokens(marker)
+        if self._estimate_text_tokens(text) <= token_budget:
+            return text
+        if token_budget <= marker_tokens:
+            return marker.strip()
+
+        low = 0
+        high = len(text)
+        best = ""
+        target_budget = token_budget - marker_tokens
+        while low <= high:
+            mid = (low + high) // 2
+            candidate = text[-mid:] if mid else ""
+            if self._estimate_text_tokens(candidate) <= target_budget:
+                best = candidate
+                low = mid + 1
+            else:
+                high = mid - 1
+        result = f"{marker}{best}"
+        while result and self._estimate_text_tokens(result) > token_budget:
+            result = result[:-1]
+        return result
+
+    def _build_chats_context(
+        self,
+        chats: list[str],
+        token_budget: int,
+    ) -> tuple[str, int, int]:
+        separator = "\n---\n"
+        separator_tokens = self._estimate_text_tokens(separator)
+        selected: list[str] = []
+        total_tokens = 0
+
+        for chat in reversed(chats):
+            chat_tokens = self._estimate_text_tokens(chat)
+            extra_tokens = chat_tokens + (separator_tokens if selected else 0)
+            if selected and total_tokens + extra_tokens > token_budget:
+                break
+            if not selected and chat_tokens > token_budget:
+                trimmed = self._trim_text_to_token_budget(chat, token_budget)
+                return trimmed, len(chats) - 1, self._estimate_text_tokens(trimmed)
+            selected.append(chat)
+            total_tokens += extra_tokens
+
+        selected.reverse()
+        omitted = len(chats) - len(selected)
+        chats_str = separator.join(selected)
+        if omitted > 0:
+            omitted_notice = (
+                f"[{omitted} earlier group messages omitted due to token budget]"
+            )
+            chats_str = f"{omitted_notice}{separator}{chats_str}"
+            total_tokens += (
+                self._estimate_text_tokens(omitted_notice) + separator_tokens
+            )
+        if total_tokens > token_budget:
+            chats_str = self._trim_text_to_token_budget(chats_str, token_budget)
+            total_tokens = self._estimate_text_tokens(chats_str)
+        return chats_str, omitted, total_tokens
+
     async def remove_session(self, event: AstrMessageEvent) -> int:
         cnt = 0
         if event.unified_msg_origin in self.session_chats:
@@ -125,6 +210,11 @@ async def handle_message(self, event: AstrMessageEvent) -> None:
                     parts.append(f" {comp.text}")
                 elif isinstance(comp, Image):
                     if cfg["image_caption"]:
+                        logger.warning(
+                            "Group ICL image caption is enabled. Each group image may trigger an extra multimodal request. umo=%s, provider=%s",
+                            event.unified_msg_origin,
+                            cfg["image_caption_provider_id"],
+                        )
                         try:
                             url = comp.url if comp.url else comp.file
                             if not url:
@@ -153,9 +243,19 @@ async def on_req_llm(self, event: AstrMessageEvent, req: ProviderRequest) -> Non
         if event.unified_msg_origin not in self.session_chats:
             return
 
-        chats_str = "\n---\n".join(self.session_chats[event.unified_msg_origin])
-
         cfg = self.cfg(event)
+        chats_str, omitted, estimated_tokens = self._build_chats_context(
+            self.session_chats[event.unified_msg_origin],
+            cfg["group_icl_token_budget"],
+        )
+        if omitted > 0:
+            logger.warning(
+                "Group ICL context truncated by token budget. umo=%s, omitted=%s, estimated_tokens=%s, budget=%s",
+                event.unified_msg_origin,
+                omitted,
+                estimated_tokens,
+                cfg["group_icl_token_budget"],
+            )
         if cfg["enable_active_reply"]:
             prompt = req.prompt
             req.prompt = (
@@ -166,10 +266,17 @@ async def on_req_llm(self, event: AstrMessageEvent, req: ProviderRequest) -> Non
             )
             req.contexts = []  # 清空上下文，当使用了主动回复，所有聊天记录都在一个prompt中。
         else:
-            req.system_prompt += (
-                "You are now in a chatroom. The chat history is as follows: \n"
+            req.extra_user_content_parts.append(
+                TextPart(
+                    text=(
+                        "Use the following recent group chat context only as background "
+                        "for this request.\n"
+                        "[Group Chat Context]\n"
+                        "Recent group chat messages, newest messages are kept when truncated:\n"
+                        f"{chats_str}"
+                    )
+                ).mark_as_temp()
             )
-            req.system_prompt += chats_str
 
     async def after_req_llm(
         self, event: AstrMessageEvent, llm_resp: LLMResponse

astrbot/core/agent/runners/tool_loop_agent_runner.py

@@ -110,6 +110,8 @@ class _ToolExecutionInterrupted(Exception):
 class ToolLoopAgentRunner(BaseAgentRunner[TContext]):
     TOOL_RESULT_MAX_ESTIMATED_TOKENS = 27_500
     TOOL_RESULT_PREVIEW_MAX_ESTIMATED_TOKENS = 7000
+    REQUEST_WARN_ESTIMATED_INPUT_TOKENS = 16_000
+    REQUEST_WARN_IMAGE_COUNT = 1
     EMPTY_OUTPUT_RETRY_ATTEMPTS = 3
     EMPTY_OUTPUT_RETRY_WAIT_MIN_S = 1
     EMPTY_OUTPUT_RETRY_WAIT_MAX_S = 4
@@ -176,6 +178,43 @@ def _get_persona_custom_error_message(self) -> str | None:
         event = getattr(self.run_context.context, "event", None)
         return extract_persona_custom_error_message_from_event(event)
 
+    @staticmethod
+    def _count_image_parts(messages: list[Message]) -> int:
+        count = 0
+        for message in messages:
+            if isinstance(message.content, list):
+                count += sum(
+                    1 for part in message.content if isinstance(part, ImageURLPart)
+                )
+        return count
+
+    def _log_request_cost_preflight(self) -> None:
+        estimated_input_tokens = EstimateTokenCounter().count_tokens(
+            self.run_context.messages
+        )
+        image_count = self._count_image_parts(self.run_context.messages)
+        logger.debug(
+            "LLM request preflight. provider=%s, model=%s, estimated_input_tokens=%s, image_count=%s",
+            self.provider.provider_config.get("id", ""),
+            self.provider.get_model(),
+            estimated_input_tokens,
+            image_count,
+        )
+        if estimated_input_tokens >= self.REQUEST_WARN_ESTIMATED_INPUT_TOKENS:
+            logger.warning(
+                "LLM request has high estimated input tokens. provider=%s, model=%s, estimated_input_tokens=%s",
+                self.provider.provider_config.get("id", ""),
+                self.provider.get_model(),
+                estimated_input_tokens,
+            )
+        if image_count > self.REQUEST_WARN_IMAGE_COUNT:
+            logger.warning(
+                "LLM request contains multiple images. provider=%s, model=%s, image_count=%s",
+                self.provider.provider_config.get("id", ""),
+                self.provider.get_model(),
+                image_count,
+            )
+
     async def _complete_with_assistant_response(self, llm_resp: LLMResponse) -> None:
         """Finalize the current step as a plain assistant response with no tool calls."""
         self.final_llm_resp = llm_resp
@@ -711,6 +750,7 @@ async def step(self):
             self.run_context.messages, trusted_token_usage=token_usage
         )
         self._simple_print_message_role("[AftCompact]")
+        self._log_request_cost_preflight()
 
         async for llm_response in self._iter_llm_responses_with_fallback():
             if llm_response.is_chunk:

astrbot/core/astr_main_agent.py

@@ -152,10 +152,10 @@ class MainAgentBuildConfig:
     """The number of most recent turns to keep during llm_compress strategy."""
     llm_compress_provider_id: str = ""
     """The provider ID for the LLM used in context compression."""
-    max_context_length: int = -1
+    max_context_length: int = 30
     """The maximum number of turns to keep in context. -1 means no limit.
     This enforce max turns before compression"""
-    dequeue_context_length: int = 1
+    dequeue_context_length: int = 10
     """The number of oldest turns to remove when context length limit is reached."""
     fallback_max_context_tokens: int = 128000
     """Fallback max context tokens. When max_context_tokens is 0 and the model is not in LLM_METADATAS, use this value."""

astrbot/core/config/default.py

@@ -130,8 +130,8 @@
         ),
         "llm_compress_keep_recent": 6,
         "llm_compress_provider_id": "",
-        "max_context_length": -1,
-        "dequeue_context_length": 1,
+        "max_context_length": 30,
+        "dequeue_context_length": 10,
         "streaming_response": False,
         "show_tool_use_status": False,
         "show_tool_call_result": False,
@@ -217,7 +217,8 @@
     },
     "provider_ltm_settings": {
         "group_icl_enable": False,
-        "group_message_max_cnt": 300,
+        "group_message_max_cnt": 50,
+        "group_icl_token_budget": 4000,
         "image_caption": False,
         "image_caption_provider_id": "",
         "active_reply": {
@@ -2862,6 +2863,9 @@
                     "group_message_max_cnt": {
                         "type": "int",
                     },
+                    "group_icl_token_budget": {
+                        "type": "int",
+                    },
                     "image_caption": {
                         "type": "bool",
                     },
@@ -4079,6 +4083,11 @@
                         "description": "最大消息数量",
                         "type": "int",
                     },
+                    "provider_ltm_settings.group_icl_token_budget": {
+                        "description": "群聊上下文 Token 预算",
+                        "type": "int",
+                        "hint": "每次 LLM 请求注入的群聊上下文近似 token 上限。降低该值可减少费用并降低缓存失效影响。",
+                    },
                     "provider_ltm_settings.image_caption": {
                         "description": "自动理解图片",
                         "type": "bool",

astrbot/core/pipeline/process_stage/method/agent_sub_stages/internal.py

@@ -240,10 +240,19 @@ async def process(
                         and not event.platform_meta.support_streaming_message
                     )
 
+                    system_prompt_before_hooks = req.system_prompt or ""
                     if await call_event_hook(event, EventType.OnLLMRequestEvent, req):
                         if reset_coro:
                             reset_coro.close()
                         return
+                    system_prompt_after_hooks = req.system_prompt or ""
+                    if system_prompt_after_hooks != system_prompt_before_hooks:
+                        logger.warning(
+                            "LLM system prompt was modified by request hooks. umo=%s, before_chars=%s, after_chars=%s",
+                            event.unified_msg_origin,
+                            len(system_prompt_before_hooks),
+                            len(system_prompt_after_hooks),
+                        )
 
                     # apply reset
                     if reset_coro:

tests/unit/test_long_term_memory_cost_safety.py

@@ -0,0 +1,85 @@
+import pytest
+
+from astrbot.api.provider import ProviderRequest
+from astrbot.builtin_stars.astrbot.long_term_memory import LongTermMemory
+from astrbot.core.agent.message import (
+    Message,
+    TextPart,
+    dump_messages_with_checkpoints,
+)
+
+
+class DummyContext:
+    def get_config(self, umo=None):
+        return {
+            "provider_settings": {
+                "image_caption_prompt": "Describe the image.",
+            },
+            "provider_ltm_settings": {
+                "group_message_max_cnt": 50,
+                "group_icl_token_budget": 30,
+                "image_caption": False,
+                "image_caption_provider_id": "",
+                "active_reply": {
+                    "enable": False,
+                    "method": "possibility_reply",
+                    "possibility_reply": 0.1,
+                    "whitelist": [],
+                },
+            },
+        }
+
+
+class DummyEvent:
+    unified_msg_origin = "group:test"
+
+
+@pytest.mark.asyncio
+async def test_group_icl_uses_user_context_part_instead_of_system_prompt():
+    ltm = LongTermMemory(None, DummyContext())
+    ltm.session_chats[DummyEvent.unified_msg_origin] = [
+        "[alice/10:00:00]: old message",
+        "[bob/10:01:00]: recent message",
+    ]
+    req = ProviderRequest(prompt="hello", system_prompt="base system")
+
+    await ltm.on_req_llm(DummyEvent(), req)
+
+    assert "old message" not in req.system_prompt
+    assert "recent message" not in req.system_prompt
+    assert len(req.extra_user_content_parts) == 1
+    part = req.extra_user_content_parts[0]
+    assert isinstance(part, TextPart)
+    assert part._no_save is True
+    assert "only as background" in part.text
+    assert "[Group Chat Context]" in part.text
+    assert "recent message" in part.text
+
+
+@pytest.mark.asyncio
+async def test_group_icl_context_is_not_persisted_in_history():
+    ltm = LongTermMemory(None, DummyContext())
+    ltm.session_chats[DummyEvent.unified_msg_origin] = [
+        "[bob/10:01:00]: recent message",
+    ]
+    req = ProviderRequest(prompt="hello", system_prompt="base system")
+
+    await ltm.on_req_llm(DummyEvent(), req)
+    message = Message.model_validate(await req.assemble_context())
+    dumped = dump_messages_with_checkpoints([message])
+
+    assert "hello" in str(dumped)
+    assert "[Group Chat Context]" not in str(dumped)
+    assert "recent message" not in str(dumped)
+
+
+def test_group_icl_context_respects_token_budget():
+    ltm = LongTermMemory(None, DummyContext())
+    chats = [f"[user{i}/10:00:0{i}]: " + ("x" * 80) for i in range(5)]
+    budget = 30
+
+    chats_str, omitted, estimated_tokens = ltm._build_chats_context(chats, budget)
+
+    assert omitted > 0
+    assert chats_str
+    assert estimated_tokens <= budget