Merge branch 'feat_qianfan' of github.com:better629/MetaGPT into feat_qianfan

metagpt/actions/research.py

@@ -133,7 +133,7 @@ class CollectLinks(Action):
                 if len(remove) == 0:
                     break
 
-        model_name = config.get_openai_llm().model
+        model_name = config.llm.model
         prompt = reduce_message_length(gen_msg(), model_name, system_text, 4096)
         logger.debug(prompt)
         queries = await self._aask(prompt, [system_text])

metagpt/provider/general_api_requestor.py

@@ -60,7 +60,8 @@ class GeneralAPIRequestor(APIRequestor):
         self, result: requests.Response, stream: bool
     ) -> Tuple[Union[bytes, Iterator[Generator]], bytes]:
         """Returns the response(s) and a bool indicating whether it is a stream."""
-        if stream and "text/event-stream" in result.headers.get("Content-Type", ""):
+        content_type = result.headers.get("Content-Type", "")
+        if stream and ("text/event-stream" in content_type or "application/x-ndjson" in content_type):
             return (
                 self._interpret_response_line(line, result.status_code, result.headers, stream=True)
                 for line in parse_stream(result.iter_lines())

metagpt/provider/openai_api.py

@@ -233,14 +233,16 @@ class OpenAILLM(BaseLLM):
             usage.prompt_tokens = count_message_tokens(messages, self.model)
             usage.completion_tokens = count_string_tokens(rsp, self.model)
         except Exception as e:
-            logger.error(f"usage calculation failed: {e}")
+            logger.warning(f"usage calculation failed: {e}")
 
         return usage
 
     def _get_max_tokens(self, messages: list[dict]):
         if not self.auto_max_tokens:
             return self.config.max_token
-        return get_max_completion_tokens(messages, self.model, self.config.max_tokens)
+        # FIXME
+        # https://community.openai.com/t/why-is-gpt-3-5-turbo-1106-max-tokens-limited-to-4096/494973/3
+        return min(get_max_completion_tokens(messages, self.model, self.config.max_token), 4096)
 
     @handle_exception
     async def amoderation(self, content: Union[str, list[str]]):

metagpt/provider/zhipuai_api.py

@@ -3,9 +3,8 @@
 # @Desc   : zhipuai LLM from https://open.bigmodel.cn/dev/api#sdk
 
 from enum import Enum
+from typing import Optional
 
-import openai
-import zhipuai
 from requests import ConnectionError
 from tenacity import (
     after_log,
@@ -14,6 +13,7 @@ from tenacity import (
     stop_after_attempt,
     wait_random_exponential,
 )
+from zhipuai.types.chat.chat_completion import Completion
 
 from metagpt.configs.llm_config import LLMConfig, LLMType
 from metagpt.logs import log_llm_stream, logger
@@ -21,6 +21,7 @@ from metagpt.provider.base_llm import BaseLLM
 from metagpt.provider.llm_provider_registry import register_provider
 from metagpt.provider.openai_api import log_and_reraise
 from metagpt.provider.zhipuai.zhipu_model_api import ZhiPuModelAPI
+from metagpt.utils.cost_manager import CostManager
 
 
 class ZhiPuEvent(Enum):
@@ -38,27 +39,22 @@ class ZhiPuAILLM(BaseLLM):
     """
 
     def __init__(self, config: LLMConfig):
-        self.__init_zhipuai(config)
-        self.llm = ZhiPuModelAPI
-        self.model = "chatglm_turbo"  # so far only one model, just use it
-        self.use_system_prompt: bool = False  # zhipuai has no system prompt when use api
         self.config = config
+        self.__init_zhipuai()
+        self.cost_manager: Optional[CostManager] = None
 
-    def __init_zhipuai(self, config: LLMConfig):
-        assert config.api_key
-        zhipuai.api_key = config.api_key
-        # due to use openai sdk, set the api_key but it will't be used.
-        # openai.api_key = zhipuai.api_key  # due to use openai sdk, set the api_key but it will't be used.
-        if config.proxy:
-            # FIXME: openai v1.x sdk has no proxy support
-            openai.proxy = config.proxy
+    def __init_zhipuai(self):
+        assert self.config.api_key
+        self.api_key = self.config.api_key
+        self.model = self.config.model  # so far, it support glm-3-turbo、glm-4
+        self.llm = ZhiPuModelAPI(api_key=self.api_key)
 
     def _const_kwargs(self, messages: list[dict], stream: bool = False) -> dict:
         kwargs = {"model": self.model, "messages": messages, "stream": stream, "temperature": 0.3}
         return kwargs
 
     def completion(self, messages: list[dict], timeout=3) -> dict:
-        resp = self.llm.chat.completions.create(**self._const_kwargs(messages))
+        resp: Completion = self.llm.chat.completions.create(**self._const_kwargs(messages))
         usage = resp.usage.model_dump()
         self._update_costs(usage)
         return resp.model_dump()

metagpt/roles/ci/code_interpreter.py

@@ -72,11 +72,7 @@ class CodeInterpreter(Role):
                 if ReviewConst.CHANGE_WORDS[0] in review:
                     counter = 0  # redo the task again with help of human suggestions
 
-        py_code = (
-            code["code"] if code.get("language") == "python" else ""
-        )  # use python code as final code; for markdown, return the rendered result instead of the code itself
-
-        return py_code, result, success
+        return code["code"], result, success
 
     async def _write_code(self):
         todo = WriteCodeWithoutTools() if not self.use_tools else WriteCodeWithTools(selected_tools=self.tools)

metagpt/utils/cost_manager.py

@@ -42,6 +42,10 @@ class CostManager(BaseModel):
         """
         self.total_prompt_tokens += prompt_tokens
         self.total_completion_tokens += completion_tokens
+        if model not in TOKEN_COSTS:
+            logger.warning(f"Model {model} not found in TOKEN_COSTS.")
+            return
+
         cost = (
             prompt_tokens * self.token_costs[model]["prompt"]
             + completion_tokens * self.token_costs[model]["completion"]

metagpt/utils/text.py

@@ -25,7 +25,7 @@ def reduce_message_length(
     """
     max_token = TOKEN_MAX.get(model_name, 2048) - count_string_tokens(system_text, model_name) - reserved
     for msg in msgs:
-        if count_string_tokens(msg, model_name) < max_token:
+        if count_string_tokens(msg, model_name) < max_token or model_name not in TOKEN_MAX:
             return msg
 
     raise RuntimeError("fail to reduce message length")
@@ -93,7 +93,7 @@ def split_paragraph(paragraph: str, sep: str = ".,", count: int = 2) -> list[str
             continue
         ret = ["".join(j) for j in _split_by_count(sentences, count)]
         return ret
-    return _split_by_count(paragraph, count)
+    return list(_split_by_count(paragraph, count))
 
 
 def decode_unicode_escape(text: str) -> str:

metagpt/utils/token_counter.py

@@ -32,8 +32,8 @@ TOKEN_COSTS = {
     "gpt-4-vision-preview": {"prompt": 0.01, "completion": 0.03},  # TODO add extra image price calculator
     "gpt-4-1106-vision-preview": {"prompt": 0.01, "completion": 0.03},
     "text-embedding-ada-002": {"prompt": 0.0004, "completion": 0.0},
-    "glm-3-turbo": {"prompt": 0.0, "completion": 0.0007},  # 128k version, prompt + completion tokens=0.005￥/k-tokens
-    "glm-4": {"prompt": 0.0, "completion": 0.014},  # 128k version, prompt + completion tokens=0.1￥/k-tokens
+    "glm-3-turbo": {"prompt": 0.0007, "completion": 0.0007},  # 128k version, prompt + completion tokens=0.005￥/k-tokens
+    "glm-4": {"prompt": 0.014, "completion": 0.014},  # 128k version, prompt + completion tokens=0.1￥/k-tokens
     "gemini-pro": {"prompt": 0.00025, "completion": 0.0005},
 }
 
@@ -140,7 +140,8 @@ TOKEN_MAX = {
     "gpt-4-vision-preview": 128000,
     "gpt-4-1106-vision-preview": 128000,
     "text-embedding-ada-002": 8192,
-    "chatglm_turbo": 32768,
+    "glm-3-turbo": 128000,
+    "glm-4": 128000,
     "gemini-pro": 32768,
 }