Multi language support - fern debug (#1307)

---------

Co-authored-by: Louis <lpglm@orange.fr>
Co-authored-by: LeMoussel <cnhx27@gmail.com>

.github/workflows/preview-docs.yml

@@ -28,7 +28,7 @@ jobs:
         env:
           FERN_TOKEN: ${{ secrets.FERN_TOKEN }}
         run: |
-          output=$(fern generate --docs --preview)
+          output=$(fern generate --docs --preview --log-level debug)
           echo "$output"
           # Extract the URL
           preview_url=$(echo "$output" | grep -oP '(?<=Published docs to )https://[^\s]*')

fern/docs.yml

@@ -44,6 +44,10 @@ navigation:
   # Manual of privateGPT: how to use it and configure it
   - tab: manual
     layout:
+      - section: General configuration
+        contents:
+          - page: Configuration
+            path: ./docs/pages/manual/settings.mdx
       - section: Document management
         contents:
           - page: Ingestion

fern/docs/pages/manual/settings.mdx

@@ -1,4 +1,4 @@
-# Configure your private GPT
+# Settings and profiles for your private GPT
 
 The configuration of your private GPT server is done thanks to `settings` files (more precisely `settings.yaml`).
 These text files are written using the [YAML](https://en.wikipedia.org/wiki/YAML) syntax.

fern/docs/pages/recipes/list-llm.mdx

@@ -1,8 +1,41 @@
-## List of working LLM
+# List of working LLM
 
 **Do you have any working combination of LLM and embeddings?**
 Please open a PR to add it to the list, and come on our Discord to tell us about it!
 
+## Prompt style
+
+LLMs might have been trained with different prompt styles.
+The prompt style is the way the prompt is written, and how the system message is injected in the prompt.
+
+For example, `llama2` looks like this:
+```text
+<s>[INST] <<SYS>>
+{{ system_prompt }}
+<</SYS>>
+
+{{ user_message }} [/INST]
+```
+
+While `default` (the `llama_index` default) looks like this:
+```text
+system: {{ system_prompt }}
+user: {{ user_message }}
+assistant: {{ assistant_message }}
+```
+
+And the "`tag`" style looks like this:
+
+```text
+<|system|>: {{ system_prompt }}
+<|user|>: {{ user_message }}
+<|assistant|>: {{ assistant_message }}
+```
+
+Some LLMs will not understand this prompt style, and will not work (returning nothing).
+You can try to change the prompt style to `default` (or `tag`) in the settings, and it will
+change the way the messages are formatted to be passed to the LLM.
+
 ## Example of configuration
 
 You might want to change the prompt depending on the language and model you are using.
@@ -15,6 +48,7 @@ local:
   llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.1-GGUF
   llm_hf_model_file: mistral-7b-instruct-v0.1.Q4_K_M.gguf
   embedding_hf_model_name: BAAI/bge-small-en-v1.5
+  prompt_style: "llama2"
 ```
 
 ### French, with instructions
@@ -25,4 +59,26 @@ local:
   llm_hf_repo_id: TheBloke/Vigogne-2-7B-Instruct-GGUF
   llm_hf_model_file: vigogne-2-7b-instruct.Q4_K_M.gguf
   embedding_hf_model_name: dangvantuan/sentence-camembert-base
+  prompt_style: "default"
+  # prompt_style: "tag" # also works
+  # The default system prompt is injected only when the `prompt_style` != default, and there are no system message in the discussion
+  # default_system_prompt: Vous êtes un assistant IA qui répond à la question posée à la fin en utilisant le contexte suivant. Si vous ne connaissez pas la réponse, dites simplement que vous ne savez pas, n'essayez pas d'inventer une réponse. Veuillez répondre exclusivement en français.
 ```
+
+You might want to change the prompt as the one above might not directly answer your question.
+You can read online about how to write a good prompt, but in a nutshell, make it (extremely) directive.
+
+You can try and troubleshot your prompt by writing multiline requests in the UI, while
+writing your interaction with the model, for example:
+
+```text
+Tu es un programmeur senior qui programme en python et utilise le framework fastapi. Ecrit moi un serveur qui retourne "hello world".
+```
+
+Another example:
+```text
+Context: None
+Situation: tu es au milieu d'un champ.
+Tache: va a la rivière, en bas du champ.
+Décrit comment aller a la rivière.
+```

fern/fern.config.json

@@ -1,4 +1,4 @@
 {
   "organization": "privategpt",
-  "version": "0.15.0-rc80"
+  "version": "0.15.3"
 }

private_gpt/components/llm/llm_component.py

@@ -1,8 +1,8 @@
 from injector import inject, singleton
 from llama_index.llms import MockLLM
 from llama_index.llms.base import LLM
-from llama_index.llms.llama_utils import completion_to_prompt, messages_to_prompt
 
+from private_gpt.components.llm.prompt_helper import get_prompt_style
 from private_gpt.paths import models_path
 from private_gpt.settings.settings import Settings
 
@@ -17,6 +17,11 @@ class LLMComponent:
             case "local":
                 from llama_index.llms import LlamaCPP
 
+                prompt_style_cls = get_prompt_style(settings.local.prompt_style)
+                prompt_style = prompt_style_cls(
+                    default_system_prompt=settings.local.default_system_prompt
+                )
+
                 self.llm = LlamaCPP(
                     model_path=str(models_path / settings.local.llm_hf_model_file),
                     temperature=0.1,
@@ -27,8 +32,8 @@ class LLMComponent:
                     # All to GPU
                     model_kwargs={"n_gpu_layers": -1},
                     # transform inputs into Llama2 format
-                    messages_to_prompt=messages_to_prompt,
-                    completion_to_prompt=completion_to_prompt,
+                    messages_to_prompt=prompt_style.messages_to_prompt,
+                    completion_to_prompt=prompt_style.completion_to_prompt,
                     verbose=True,
                 )
 

private_gpt/components/llm/prompt_helper.py

@@ -0,0 +1,179 @@
+import abc
+import logging
+from collections.abc import Sequence
+from typing import Any, Literal
+
+from llama_index.llms import ChatMessage, MessageRole
+from llama_index.llms.llama_utils import (
+    DEFAULT_SYSTEM_PROMPT,
+    completion_to_prompt,
+    messages_to_prompt,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class AbstractPromptStyle(abc.ABC):
+    """Abstract class for prompt styles.
+
+    This class is used to format a series of messages into a prompt that can be
+    understood by the models. A series of messages represents the interaction(s)
+    between a user and an assistant. This series of messages can be considered as a
+    session between a user X and an assistant Y.This session holds, through the
+    messages, the state of the conversation. This session, to be understood by the
+    model, needs to be formatted into a prompt (i.e. a string that the models
+    can understand). Prompts can be formatted in different ways,
+    depending on the model.
+
+    The implementations of this class represent the different ways to format a
+    series of messages into a prompt.
+    """
+
+    @abc.abstractmethod
+    def __init__(self, *args: Any, **kwargs: Any) -> None:
+        logger.debug("Initializing prompt_style=%s", self.__class__.__name__)
+
+    @abc.abstractmethod
+    def _messages_to_prompt(self, messages: Sequence[ChatMessage]) -> str:
+        pass
+
+    @abc.abstractmethod
+    def _completion_to_prompt(self, completion: str) -> str:
+        pass
+
+    def messages_to_prompt(self, messages: Sequence[ChatMessage]) -> str:
+        prompt = self._messages_to_prompt(messages)
+        logger.debug("Got for messages='%s' the prompt='%s'", messages, prompt)
+        return prompt
+
+    def completion_to_prompt(self, completion: str) -> str:
+        prompt = self._completion_to_prompt(completion)
+        logger.debug("Got for completion='%s' the prompt='%s'", completion, prompt)
+        return prompt
+
+
+class AbstractPromptStyleWithSystemPrompt(AbstractPromptStyle, abc.ABC):
+    _DEFAULT_SYSTEM_PROMPT = DEFAULT_SYSTEM_PROMPT
+
+    def __init__(self, default_system_prompt: str | None) -> None:
+        super().__init__()
+        logger.debug("Got default_system_prompt='%s'", default_system_prompt)
+        self.default_system_prompt = default_system_prompt
+
+
+class DefaultPromptStyle(AbstractPromptStyle):
+    """Default prompt style that uses the defaults from llama_utils.
+
+    It basically passes None to the LLM, indicating it should use
+    the default functions.
+    """
+
+    def __init__(self, *args: Any, **kwargs: Any) -> None:
+        super().__init__(*args, **kwargs)
+
+        # Hacky way to override the functions
+        # Override the functions to be None, and pass None to the LLM.
+        self.messages_to_prompt = None  # type: ignore[method-assign, assignment]
+        self.completion_to_prompt = None  # type: ignore[method-assign, assignment]
+
+    def _messages_to_prompt(self, messages: Sequence[ChatMessage]) -> str:
+        return ""
+
+    def _completion_to_prompt(self, completion: str) -> str:
+        return ""
+
+
+class Llama2PromptStyle(AbstractPromptStyleWithSystemPrompt):
+    """Simple prompt style that just uses the default llama_utils functions.
+
+    It transforms the sequence of messages into a prompt that should look like:
+    ```text
+    <s> [INST] <<SYS>> your system prompt here. <</SYS>>
+
+    user message here [/INST] assistant (model) response here </s>
+    ```
+    """
+
+    def __init__(self, default_system_prompt: str | None = None) -> None:
+        # If no system prompt is given, the default one of the implementation is used.
+        super().__init__(default_system_prompt=default_system_prompt)
+
+    def _messages_to_prompt(self, messages: Sequence[ChatMessage]) -> str:
+        return messages_to_prompt(messages, self.default_system_prompt)
+
+    def _completion_to_prompt(self, completion: str) -> str:
+        return completion_to_prompt(completion, self.default_system_prompt)
+
+
+class TagPromptStyle(AbstractPromptStyleWithSystemPrompt):
+    """Tag prompt style (used by Vigogne) that uses the prompt style `<|ROLE|>`.
+
+    It transforms the sequence of messages into a prompt that should look like:
+    ```text
+    <|system|>: your system prompt here.
+    <|user|>: user message here
+    (possibly with context and question)
+    <|assistant|>: assistant (model) response here.
+    ```
+
+    FIXME: should we add surrounding `<s>` and `</s>` tags, like in llama2?
+    """
+
+    def __init__(self, default_system_prompt: str | None = None) -> None:
+        # We have to define a default system prompt here as the LLM will not
+        # use the default llama_utils functions.
+        default_system_prompt = default_system_prompt or self._DEFAULT_SYSTEM_PROMPT
+        super().__init__(default_system_prompt)
+        self.system_prompt: str = default_system_prompt
+
+    def _messages_to_prompt(self, messages: Sequence[ChatMessage]) -> str:
+        messages = list(messages)
+        if messages[0].role != MessageRole.SYSTEM:
+            logger.info(
+                "Adding system_promt='%s' to the given messages as there are none given in the session",
+                self.system_prompt,
+            )
+            messages = [
+                ChatMessage(content=self.system_prompt, role=MessageRole.SYSTEM),
+                *messages,
+            ]
+        return self._format_messages_to_prompt(messages)
+
+    def _completion_to_prompt(self, completion: str) -> str:
+        return (
+            f"<|system|>: {self.system_prompt.strip()}\n"
+            f"<|user|>: {completion.strip()}\n"
+            "<|assistant|>: "
+        )
+
+    @staticmethod
+    def _format_messages_to_prompt(messages: list[ChatMessage]) -> str:
+        """Format message to prompt with `<|ROLE|>: MSG` style."""
+        assert messages[0].role == MessageRole.SYSTEM
+        prompt = ""
+        for message in messages:
+            role = message.role
+            content = message.content or ""
+            message_from_user = f"<|{role.lower()}|>: {content.strip()}"
+            message_from_user += "\n"
+            prompt += message_from_user
+        # we are missing the last <|assistant|> tag that will trigger a completion
+        prompt += "<|assistant|>: "
+        return prompt
+
+
+def get_prompt_style(
+    prompt_style: Literal["default", "llama2", "tag"] | None
+) -> type[AbstractPromptStyle]:
+    """Get the prompt style to use from the given string.
+
+    :param prompt_style: The prompt style to use.
+    :return: The prompt style to use.
+    """
+    if prompt_style is None or prompt_style == "default":
+        return DefaultPromptStyle
+    elif prompt_style == "llama2":
+        return Llama2PromptStyle
+    elif prompt_style == "tag":
+        return TagPromptStyle
+    raise ValueError(f"Unknown prompt_style='{prompt_style}'")

private_gpt/settings/settings.py

@@ -91,7 +91,28 @@ class VectorstoreSettings(BaseModel):
 class LocalSettings(BaseModel):
     llm_hf_repo_id: str
     llm_hf_model_file: str
-    embedding_hf_model_name: str
+    embedding_hf_model_name: str = Field(
+        description="Name of the HuggingFace model to use for embeddings"
+    )
+    prompt_style: Literal["default", "llama2", "tag"] = Field(
+        "llama2",
+        description=(
+            "The prompt style to use for the chat engine. "
+            "If `default` - use the default prompt style from the llama_index. It should look like `role: message`.\n"
+            "If `llama2` - use the llama2 prompt style from the llama_index. Based on `<s>`, `[INST]` and `<<SYS>>`.\n"
+            "If `tag` - use the `tag` prompt style. It should look like `<|role|>: message`. \n"
+            "`llama2` is the historic behaviour. `default` might work better with your custom models."
+        ),
+    )
+    default_system_prompt: str | None = Field(
+        None,
+        description=(
+            "The default system prompt to use for the chat engine. "
+            "If none is given - use the default system prompt (from the llama_index). "
+            "Please note that the default prompt might not be the same for all prompt styles. "
+            "Also note that this is only used if the first message is not a system message. "
+        ),
+    )
 
 
 class SagemakerSettings(BaseModel):

settings.yaml

@@ -30,6 +30,7 @@ qdrant:
   path: local_data/private_gpt/qdrant
 
 local:
+  prompt_style: "llama2"
   llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.1-GGUF
   llm_hf_model_file: mistral-7b-instruct-v0.1.Q4_K_M.gguf
   embedding_hf_model_name: BAAI/bge-small-en-v1.5

tests/test_prompt_helper.py

@@ -0,0 +1,128 @@
+import pytest
+from llama_index.llms import ChatMessage, MessageRole
+
+from private_gpt.components.llm.prompt_helper import (
+    DefaultPromptStyle,
+    Llama2PromptStyle,
+    TagPromptStyle,
+    get_prompt_style,
+)
+
+
+@pytest.mark.parametrize(
+    ("prompt_style", "expected_prompt_style"),
+    [
+        ("default", DefaultPromptStyle),
+        ("llama2", Llama2PromptStyle),
+        ("tag", TagPromptStyle),
+    ],
+)
+def test_get_prompt_style_success(prompt_style, expected_prompt_style):
+    assert get_prompt_style(prompt_style) == expected_prompt_style
+
+
+def test_get_prompt_style_failure():
+    prompt_style = "unknown"
+    with pytest.raises(ValueError) as exc_info:
+        get_prompt_style(prompt_style)
+    assert str(exc_info.value) == f"Unknown prompt_style='{prompt_style}'"
+
+
+def test_tag_prompt_style_format():
+    prompt_style = TagPromptStyle()
+    messages = [
+        ChatMessage(content="You are an AI assistant.", role=MessageRole.SYSTEM),
+        ChatMessage(content="Hello, how are you doing?", role=MessageRole.USER),
+    ]
+
+    expected_prompt = (
+        "<|system|>: You are an AI assistant.\n"
+        "<|user|>: Hello, how are you doing?\n"
+        "<|assistant|>: "
+    )
+
+    assert prompt_style.messages_to_prompt(messages) == expected_prompt
+
+
+def test_tag_prompt_style_format_with_system_prompt():
+    system_prompt = "This is a system prompt from configuration."
+    prompt_style = TagPromptStyle(default_system_prompt=system_prompt)
+    messages = [
+        ChatMessage(content="Hello, how are you doing?", role=MessageRole.USER),
+    ]
+
+    expected_prompt = (
+        f"<|system|>: {system_prompt}\n"
+        "<|user|>: Hello, how are you doing?\n"
+        "<|assistant|>: "
+    )
+
+    assert prompt_style.messages_to_prompt(messages) == expected_prompt
+
+    messages = [
+        ChatMessage(
+            content="FOO BAR Custom sys prompt from messages.", role=MessageRole.SYSTEM
+        ),
+        ChatMessage(content="Hello, how are you doing?", role=MessageRole.USER),
+    ]
+
+    expected_prompt = (
+        "<|system|>: FOO BAR Custom sys prompt from messages.\n"
+        "<|user|>: Hello, how are you doing?\n"
+        "<|assistant|>: "
+    )
+
+    assert prompt_style.messages_to_prompt(messages) == expected_prompt
+
+
+def test_llama2_prompt_style_format():
+    prompt_style = Llama2PromptStyle()
+    messages = [
+        ChatMessage(content="You are an AI assistant.", role=MessageRole.SYSTEM),
+        ChatMessage(content="Hello, how are you doing?", role=MessageRole.USER),
+    ]
+
+    expected_prompt = (
+        "<s> [INST] <<SYS>>\n"
+        " You are an AI assistant. \n"
+        "<</SYS>>\n"
+        "\n"
+        " Hello, how are you doing? [/INST]"
+    )
+
+    assert prompt_style.messages_to_prompt(messages) == expected_prompt
+
+
+def test_llama2_prompt_style_with_system_prompt():
+    system_prompt = "This is a system prompt from configuration."
+    prompt_style = Llama2PromptStyle(default_system_prompt=system_prompt)
+    messages = [
+        ChatMessage(content="Hello, how are you doing?", role=MessageRole.USER),
+    ]
+
+    expected_prompt = (
+        "<s> [INST] <<SYS>>\n"
+        f" {system_prompt} \n"
+        "<</SYS>>\n"
+        "\n"
+        " Hello, how are you doing? [/INST]"
+    )
+
+    assert prompt_style.messages_to_prompt(messages) == expected_prompt
+
+    messages = [
+        ChatMessage(
+            content="FOO BAR Custom sys prompt from messages.", role=MessageRole.SYSTEM
+        ),
+        ChatMessage(content="Hello, how are you doing?", role=MessageRole.USER),
+    ]
+
+    expected_prompt = (
+        "<s> [INST] <<SYS>>\n"
+        " FOO BAR Custom sys prompt from messages. \n"
+        "<</SYS>>\n"
+        "\n"
+        " Hello, how are you doing? [/INST]"
+    )
+
+    assert prompt_style.messages_to_prompt(messages) == expected_prompt