Expose templates for customisation in providers (jupyterlab#581)

* Expose chat template for customisation in providers * Enable completion template customization, including suffix * Remove `inline` from "inline completion" as per review request
dbelgrod · Jan 25, 2024 · 7bf3789 · 7bf3789
1 parent 205f552
commit 7bf3789
Show file tree

Hide file tree

Showing 3 changed files with 105 additions and 86 deletions.
diff --git a/packages/jupyter-ai-magics/jupyter_ai_magics/providers.py b/packages/jupyter-ai-magics/jupyter_ai_magics/providers.py
@@ -11,7 +11,13 @@
 from langchain.chat_models.base import BaseChatModel
 from langchain.llms.sagemaker_endpoint import LLMContentHandler
 from langchain.llms.utils import enforce_stop_tokens
-from langchain.prompts import PromptTemplate
+from langchain.prompts import (
+    ChatPromptTemplate,
+    HumanMessagePromptTemplate,
+    MessagesPlaceholder,
+    PromptTemplate,
+    SystemMessagePromptTemplate,
+)
 from langchain.pydantic_v1 import BaseModel, Extra, root_validator
 from langchain.schema import LLMResult
 from langchain.utils import get_from_dict_or_env
@@ -42,6 +48,49 @@
     from pydantic.main import ModelMetaclass
 
 
+CHAT_SYSTEM_PROMPT = """
+You are Jupyternaut, a conversational assistant living in JupyterLab to help users.
+You are not a language model, but rather an application built on a foundation model from {provider_name} called {local_model_id}.
+You are talkative and you provide lots of specific details from the foundation model's context.
+You may use Markdown to format your response.
+Code blocks must be formatted in Markdown.
+Math should be rendered with inline TeX markup, surrounded by $.
+If you do not know the answer to a question, answer truthfully by responding that you do not know.
+The following is a friendly conversation between you and a human.
+""".strip()
+
+CHAT_DEFAULT_TEMPLATE = """Current conversation:
+{history}
+Human: {input}
+AI:"""
+
+
+COMPLETION_SYSTEM_PROMPT = """
+You are an application built to provide helpful code completion suggestions.
+You should only produce code. Keep comments to minimum, use the
+programming language comment syntax. Produce clean code.
+The code is written in JupyterLab, a data analysis and code development
+environment which can execute code extended with additional syntax for
+interactive features, such as magics.
+""".strip()
+
+# only add the suffix bit if present to save input tokens/computation time
+COMPLETION_DEFAULT_TEMPLATE = """
+The document is called `{{filename}}` and written in {{language}}.
+{% if suffix %}
+The code after the completion request is:
+
+```
+{{suffix}}
+```
+{% endif %}
+
+Complete the following code:
+
+```
+{{prefix}}"""
+
+
 class EnvAuthStrategy(BaseModel):
     """Require one auth token via an environment variable."""
 
@@ -265,6 +314,55 @@ def get_prompt_template(self, format) -> PromptTemplate:
         else:
             return self.prompt_templates["text"]  # Default to plain format
 
+    def get_chat_prompt_template(self) -> PromptTemplate:
+        """
+        Produce a prompt template optimised for chat conversation.
+        The template should take two variables: history and input.
+        """
+        name = self.__class__.name
+        if self.is_chat_provider:
+            return ChatPromptTemplate.from_messages(
+                [
+                    SystemMessagePromptTemplate.from_template(
+                        CHAT_SYSTEM_PROMPT
+                    ).format(provider_name=name, local_model_id=self.model_id),
+                    MessagesPlaceholder(variable_name="history"),
+                    HumanMessagePromptTemplate.from_template("{input}"),
+                ]
+            )
+        else:
+            return PromptTemplate(
+                input_variables=["history", "input"],
+                template=CHAT_SYSTEM_PROMPT.format(
+                    provider_name=name, local_model_id=self.model_id
+                )
+                + "\n\n"
+                + CHAT_DEFAULT_TEMPLATE,
+            )
+
+    def get_completion_prompt_template(self) -> PromptTemplate:
+        """
+        Produce a prompt template optimised for inline code or text completion.
+        The template should take variables: prefix, suffix, language, filename.
+        """
+        if self.is_chat_provider:
+            return ChatPromptTemplate.from_messages(
+                [
+                    SystemMessagePromptTemplate.from_template(COMPLETION_SYSTEM_PROMPT),
+                    HumanMessagePromptTemplate.from_template(
+                        COMPLETION_DEFAULT_TEMPLATE, template_format="jinja2"
+                    ),
+                ]
+            )
+        else:
+            return PromptTemplate(
+                input_variables=["prefix", "suffix", "language", "filename"],
+                template=COMPLETION_SYSTEM_PROMPT
+                + "\n\n"
+                + COMPLETION_DEFAULT_TEMPLATE,
+                template_format="jinja2",
+            )
+
     @property
     def is_chat_provider(self):
         return isinstance(self, BaseChatModel)

diff --git a/packages/jupyter-ai/jupyter_ai/chat_handlers/default.py b/packages/jupyter-ai/jupyter_ai/chat_handlers/default.py
@@ -4,32 +4,9 @@
 from jupyter_ai_magics.providers import BaseProvider
 from langchain.chains import ConversationChain
 from langchain.memory import ConversationBufferWindowMemory
-from langchain.prompts import (
-    ChatPromptTemplate,
-    HumanMessagePromptTemplate,
-    MessagesPlaceholder,
-    PromptTemplate,
-    SystemMessagePromptTemplate,
-)
 
 from .base import BaseChatHandler, SlashCommandRoutingType
 
-SYSTEM_PROMPT = """
-You are Jupyternaut, a conversational assistant living in JupyterLab to help users.
-You are not a language model, but rather an application built on a foundation model from {provider_name} called {local_model_id}.
-You are talkative and you provide lots of specific details from the foundation model's context.
-You may use Markdown to format your response.
-Code blocks must be formatted in Markdown.
-Math should be rendered with inline TeX markup, surrounded by $.
-If you do not know the answer to a question, answer truthfully by responding that you do not know.
-The following is a friendly conversation between you and a human.
-""".strip()
-
-DEFAULT_TEMPLATE = """Current conversation:
-{history}
-Human: {input}
-AI:"""
-
 
 class DefaultChatHandler(BaseChatHandler):
     id = "default"
@@ -49,27 +26,10 @@ def create_llm_chain(
         model_parameters = self.get_model_parameters(provider, provider_params)
         llm = provider(**provider_params, **model_parameters)
 
-        if llm.is_chat_provider:
-            prompt_template = ChatPromptTemplate.from_messages(
-                [
-                    SystemMessagePromptTemplate.from_template(SYSTEM_PROMPT).format(
-                        provider_name=provider.name, local_model_id=llm.model_id
-                    ),
-                    MessagesPlaceholder(variable_name="history"),
-                    HumanMessagePromptTemplate.from_template("{input}"),
-                ]
-            )
-            self.memory = ConversationBufferWindowMemory(return_messages=True, k=2)
-        else:
-            prompt_template = PromptTemplate(
-                input_variables=["history", "input"],
-                template=SYSTEM_PROMPT.format(
-                    provider_name=provider.name, local_model_id=llm.model_id
-                )
-                + "\n\n"
-                + DEFAULT_TEMPLATE,
-            )
-            self.memory = ConversationBufferWindowMemory(k=2)
+        prompt_template = llm.get_chat_prompt_template()
+        self.memory = ConversationBufferWindowMemory(
+            return_messages=llm.is_chat_provider, k=2
+        )
 
         self.llm = llm
         self.llm_chain = ConversationChain(

diff --git a/packages/jupyter-ai/jupyter_ai/completions/handlers/default.py b/packages/jupyter-ai/jupyter_ai/completions/handlers/default.py
@@ -18,32 +18,6 @@
 )
 from .base import BaseInlineCompletionHandler
 
-SYSTEM_PROMPT = """
-You are an application built to provide helpful code completion suggestions.
-You should only produce code. Keep comments to minimum, use the
-programming language comment syntax. Produce clean code.
-The code is written in JupyterLab, a data analysis and code development
-environment which can execute code extended with additional syntax for
-interactive features, such as magics.
-""".strip()
-
-AFTER_TEMPLATE = """
-The code after the completion request is:
-
-```
-{suffix}
-```
-""".strip()
-
-DEFAULT_TEMPLATE = """
-The document is called `{filename}` and written in {language}.
-{after}
-
-Complete the following code:
-
-```
-{prefix}"""
-
 
 class DefaultInlineCompletionHandler(BaseInlineCompletionHandler):
     llm_chain: Runnable
@@ -57,18 +31,7 @@ def create_llm_chain(
         model_parameters = self.get_model_parameters(provider, provider_params)
         llm = provider(**provider_params, **model_parameters)
 
-        if llm.is_chat_provider:
-            prompt_template = ChatPromptTemplate.from_messages(
-                [
-                    SystemMessagePromptTemplate.from_template(SYSTEM_PROMPT),
-                    HumanMessagePromptTemplate.from_template(DEFAULT_TEMPLATE),
-                ]
-            )
-        else:
-            prompt_template = PromptTemplate(
-                input_variables=["prefix", "suffix", "language", "filename"],
-                template=SYSTEM_PROMPT + "\n\n" + DEFAULT_TEMPLATE,
-            )
+        prompt_template = llm.get_completion_prompt_template()
 
         self.llm = llm
         self.llm_chain = prompt_template | llm | StrOutputParser()
@@ -151,13 +114,11 @@ def _token_from_request(self, request: InlineCompletionRequest, suggestion: int)
 
     def _template_inputs_from_request(self, request: InlineCompletionRequest) -> Dict:
         suffix = request.suffix.strip()
-        # only add the suffix template if the suffix is there to save input tokens/computation time
-        after = AFTER_TEMPLATE.format(suffix=suffix) if suffix else ""
         filename = request.path.split("/")[-1] if request.path else "untitled"
 
         return {
             "prefix": request.prefix,
-            "after": after,
+            "suffix": suffix,
             "language": request.language,
             "filename": filename,
             "stop": ["\n```"],