diff --git a/docs/sphinx_doc/en/source/tutorial/206-prompt.md b/docs/sphinx_doc/en/source/tutorial/206-prompt.md index d33a8612e..f99c86024 100644 --- a/docs/sphinx_doc/en/source/tutorial/206-prompt.md +++ b/docs/sphinx_doc/en/source/tutorial/206-prompt.md @@ -241,11 +241,11 @@ messages as input. The message must obey the following rules (updated in #### Prompt Strategy -Given a list of messages, we will parse each message as follows: - -- `Msg`: Fill the `role` and `content` fields directly. If it has an `url` - field, which refers to an image, we will add it to the message. -- `List`: Parse each element in the list according to the above rules. +- If the role field of the first input message is `"system"`, +it will be treated as system prompt and the other messages will consist +dialogue history in the system message prefixed by "## Dialogue History". +- If the `url` attribute of messages is not `None`, we will gather all urls in +the `"images"` field in the returned dictionary. ```python from agentscope.models import OllamaChatWrapper @@ -268,9 +268,11 @@ print(prompt) ```bash [ - {"role": "system", "content": "You are a helpful assistant"}, - {"role": "assistant", "content": "Hi."}, - {"role": "assistant", "content": "Nice to meet you!", "images": ["https://example.com/image.jpg"]}, + { + "role": "system", + "content": "You are a helpful assistant\n\n## Dialogue History\nBob: Hi.\nAlice: Nice to meet you!", + "images": ["https://example.com/image.jpg"] + }, ] ``` diff --git a/docs/sphinx_doc/zh_CN/source/tutorial/206-prompt.md b/docs/sphinx_doc/zh_CN/source/tutorial/206-prompt.md index 5648d2193..d77fb673a 100644 --- a/docs/sphinx_doc/zh_CN/source/tutorial/206-prompt.md +++ b/docs/sphinx_doc/zh_CN/source/tutorial/206-prompt.md @@ -211,8 +211,10 @@ print(prompt) 给定一个消息列表，我们将按照以下规则解析每个消息： -- `Msg`：直接填充`role`和`content`字段。如果它有一个`url`字段，指向一个图片，我们将把它添加到消息中。 -- `List`：根据上述规则解析列表中的每个元素。 +- 如果输入的第一条信息的`role`字段是`"system"`，该条信息将被视为系统提示（system + prompt），其他信息将一起组成对话历史。对话历史将添加`"## Dialogue History"`的前缀，并与 +系统提示一起组成一条`role`为`"system"`的信息。 +- 如果输入信息中的`url`字段不为`None`，则这些url将一起被置于`"images"`对应的键值中。 ```python from agentscope.models import OllamaChatWrapper @@ -235,9 +237,11 @@ print(prompt) ```bash [ - {"role": "system", "content": "You are a helpful assistant"}, - {"role": "assistant", "content": "Hi."}, - {"role": "assistant", "content": "Nice to meet you!", "images": ["https://example.com/image.jpg"]}, + { + "role": "system", + "content": "You are a helpful assistant\n\n## Dialogue History\nBob: Hi.\nAlice: Nice to meet you!", + "images": ["https://example.com/image.jpg"] + }, ] ``` diff --git a/src/agentscope/models/ollama_model.py b/src/agentscope/models/ollama_model.py index 31b136dcb..1c38d3b8b 100644 --- a/src/agentscope/models/ollama_model.py +++ b/src/agentscope/models/ollama_model.py @@ -3,8 +3,6 @@ from abc import ABC from typing import Sequence, Any, Optional, List, Union -from loguru import logger - from agentscope.message import MessageBase from agentscope.models import ModelWrapperBase, ModelResponse from agentscope.utils.tools import _convert_to_str @@ -170,10 +168,43 @@ def format( self, *args: Union[MessageBase, Sequence[MessageBase]], ) -> List[dict]: - """A basic strategy to format the input into the required format of - Ollama Chat API. + """Format the messages for ollama Chat API. + + All messages will be formatted into a single system message with + system prompt and dialogue history. + + Note: + 1. This strategy maybe not suitable for all scenarios, + and developers are encouraged to implement their own prompt + engineering strategies. + 2. For ollama chat api, the content field shouldn't be empty string. + + Example: + + .. code-block:: python + + prompt = model.format( + Msg("system", "You're a helpful assistant", role="system"), + Msg("Bob", "Hi, how can I help you?", role="assistant"), + Msg("user", "What's the date today?", role="user") + ) + + The prompt will be as follows: + + .. code-block:: python + + [ + { + "role": "user", + "content": ( + "You're a helpful assistant\\n\\n" + "## Dialogue History\\n" + "Bob: Hi, how can I help you?\\n" + "user: What's the date today?" + ) + } + ] - Note for ollama chat api, the content field shouldn't be empty string. Args: args (`Union[MessageBase, Sequence[MessageBase]]`): @@ -185,39 +216,65 @@ def format( `List[dict]`: The formatted messages. """ - ollama_msgs = [] - for msg in args: - if msg is None: - continue - if isinstance(msg, MessageBase): - # content shouldn't be empty string - if msg.content == "": - logger.warning( - "In ollama chat API, the content field cannot be " - "empty string. To avoid error, the empty string is " - "replaced by a blank space automatically, but the " - "model may not work as expected.", - ) - msg.content = " " - - ollama_msg = { - "role": msg.role, - "content": _convert_to_str(msg.content), - } - - # image url - if msg.url is not None: - ollama_msg["images"] = [msg.url] - ollama_msgs.append(ollama_msg) - elif isinstance(msg, list): - ollama_msgs.extend(self.format(*msg)) + # Parse all information into a list of messages + input_msgs = [] + for _ in args: + if _ is None: + continue + if isinstance(_, MessageBase): + input_msgs.append(_) + elif isinstance(_, list) and all( + isinstance(__, MessageBase) for __ in _ + ): + input_msgs.extend(_) else: raise TypeError( - f"Invalid message type: {type(msg)}, `Msg` is expected.", + f"The input should be a Msg object or a list " + f"of Msg objects, got {type(_)}.", ) - return ollama_msgs + # record dialog history as a list of strings + system_content_template = [] + dialogue = [] + # TODO: here we default the url links to images + images = [] + for i, unit in enumerate(input_msgs): + if i == 0 and unit.role == "system": + # system prompt + system_prompt = _convert_to_str(unit.content) + if not system_prompt.endswith("\n"): + system_prompt += "\n" + system_content_template.append(system_prompt) + else: + # Merge all messages into a dialogue history prompt + dialogue.append( + f"{unit.name}: {_convert_to_str(unit.content)}", + ) + + if unit.url is not None: + images.append(unit.url) + + if len(dialogue) != 0: + system_content_template.extend( + ["## Dialogue History", "{dialogue_history}"], + ) + + dialogue_history = "\n".join(dialogue) + + system_content_template = "\n".join(system_content_template) + + system_message = { + "role": "system", + "content": system_content_template.format( + dialogue_history=dialogue_history, + ), + } + + if len(images) != 0: + system_message["images"] = images + + return [system_message] class OllamaEmbeddingWrapper(OllamaWrapperBase): diff --git a/tests/format_test.py b/tests/format_test.py index 3226deab4..42df7b960 100644 --- a/tests/format_test.py +++ b/tests/format_test.py @@ -84,9 +84,16 @@ def test_ollama_chat(self) -> None: # correct format ground_truth = [ - {"role": "system", "content": "You are a helpful assistant"}, - {"role": "user", "content": "What is the weather today?"}, - {"role": "assistant", "content": "It is sunny today"}, + { + "role": "system", + "content": ( + "You are a helpful assistant\n" + "\n" + "## Dialogue History\n" + "user: What is the weather today?\n" + "assistant: It is sunny today" + ), + }, ] prompt = model.format(*self.inputs) # type: ignore[arg-type] self.assertEqual(prompt, ground_truth)