Track and format chat history

Given the multiplicity of formats, formatting the prompt for chat
workflows with open models can be a real hassle and is error-prone. In
this PR we introduce a `Chat` class that allows users to track the
conversation and easily print the corresponding prompt.

Author: rlouf

docs/reference/chat.md

@@ -0,0 +1,20 @@
+# Chat history
+
+## Filter message
+
+In some situation you may want to filter the messages before building the prompt, for instance to use RAG. In this case you can subclass `Chat` and override the `filter` method:
+
+
+```python
+from prompts import Chat
+
+class RAGChat(Chat):
+
+    def filter(self):
+        filtered_message = []
+        for message in filtered_message:
+           if message.role == "user"  and "Hi" in message.content:
+               filtered_message.append(message)
+
+        return filtered_messages
+```

mkdocs.yml

@@ -76,3 +76,4 @@ nav:
       - Prompt template: reference/template.md
       - Dispatch: reference/dispatch.md
       - Special tokens: reference/special_tokens.md
+      - Chat History: reference/chat.md

prompts/chat.py

@@ -0,0 +1,81 @@
+from dataclasses import dataclass
+from enum import Enum
+from typing import List, Optional
+
+from pydantic import BaseModel
+from typing_extensions import TypedDict
+
+
+class Document(TypedDict):
+    title: str
+    text: str
+
+
+class Role(Enum):
+    system = "system"
+    user = "user"
+    assistant = "assistant"
+
+
+@dataclass
+class Message:
+    role: Role
+    content: str
+
+
+class Chat:
+
+    def __init__(
+        self,
+        system_msg: Optional[str] = None,
+        tools: Optional[List[BaseModel]] = None,
+        documents: Optional[List[Document]] = None,
+        history: List[Message] = [],
+    ):
+        self.history = history
+        self.system = system_msg
+        self.tools = tools
+        self.documents = documents
+
+    @property
+    def trimmed_history(self):
+        return self.history
+
+    def __add__(self, other: Message):
+        history = self.history
+        history.append(other)
+        return Chat(self.system, self.tools, self.documents, history=history)
+
+    def __iadd__(self, other: Message):
+        self.history.append(other)
+
+    def __getitem__(self, key):
+        if isinstance(key, int):
+            return self.history[key]
+        else:
+            raise KeyError()
+
+    def render(self, model_name: str):
+        """Render the conversation using the model's chat template.
+
+        TODO: Do this ourselves.
+
+        Parameters
+        ----------
+        model_name
+            The name of the model whose chat template we need to use.
+
+        """
+        from transformers import AutoTokenizer
+
+        conversation = []
+        if self.system is not None:
+            conversation.append({"role": "system", "content": self.system})
+        for message in self.trimmed_history:
+            conversation.append({"role": message.role, "content": message.content})
+
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        return self.tokenizer.apply_chat_template(
+            conversation, self.tools, self.documents
+        )

pyproject.toml

@@ -4,7 +4,7 @@ version = "0.1.0"
 description = "Large Language Models prompting library"
 authors = [{name = "The Outlines developers", email = "contact@dottxt.co"}]
 requires-python = ">= 3.8"
-dependencies = ["jinja2"]
+dependencies = ["jinja2", "pydantic", "transformers"]
 
 [build-system]
 requires = ["setuptools>=45", "setuptools_scm[toml]>=6.2"]
@@ -35,5 +35,5 @@ file="README.md"
 content-type = "text/markdown"
 
 [[tool.mypy.overrides]]
-module = ["jinja2", "pytest"]
+module = ["jinja2", "pydantic", "pytest", "transformers"]
 ignore_missing_imports = true

tests/test_chat.py

@@ -0,0 +1,7 @@
+from prompts.chat import Chat, Message
+
+
+def test_simple():
+    chat = Chat("system message")
+    new_chat = chat + Message("user", "new user message")
+    new_chat += Message("assistant", "new assistant message")