Add LLM Builder UI

autollm/cli.py

@@ -0,0 +1,232 @@
+import os
+
+import gradio as gr
+import llama_index
+from llama_index import Document
+
+from autollm.auto.llm import AutoLiteLLM
+from autollm.auto.query_engine import AutoQueryEngine
+from autollm.serve.llm_utils import create_custom_llm
+from autollm.utils.document_reading import read_files_as_documents
+
+llama_index.set_global_handler("simple")
+
+DEFAULT_LLM_MODEL = "gpt-4-0125-preview"
+OPENAI_MODELS = [
+    "gpt-4-0125-preview", "gpt-4-1106-preview", "gpt-4", "gpt-4-0613", "gpt-4-32k", "gpt-4-32k-0613",
+    "gpt-3.5-turbo-1106", "gpt-3.5-turbo", "gpt-3.5-turbo-16k", "gpt-3.5-turbo-instruct"
+]
+GEMINI_MODELS = [
+    "gemini-pro",
+]
+
+
+def determine_llm_model(model_selections: list[tuple[str, bool]]) -> str:
+    """
+    Determines which LLM model is selected based on user input.
+
+    Parameters:
+    model_selections (list of tuples): List containing tuples of (model_name, input_variable).
+
+    Returns:
+    str: Selected LLM model name.
+    """
+    selected_models = []
+    for model_name, is_selected in model_selections:
+        if is_selected:
+            selected_models.append(model_name)
+
+    if len(selected_models) != 1:
+        raise ValueError("Exactly one LLM model must be selected.")
+    return selected_models[0]
+
+
+def create_app(
+        use_openai, openai_model, openai_api_key, use_gemini, gemini_model, gemini_api_key, what_to_make_area,
+        uploaded_files, webpage_input, config_file):
+    global query_engine
+    progress = gr.Progress()
+
+    os.environ["OPENAI_API_KEY"] = openai_api_key
+    os.environ["GEMINI_API_KEY"] = gemini_api_key
+
+    progress(0.2, desc="Reading files...")
+    file_documents = read_files_as_documents(input_files=uploaded_files)
+
+    progress(0.6, desc="Updating LLM...")
+    custom_llm = create_custom_llm(user_prompt=what_to_make_area, config=config_file)
+    emoji, name, description, instructions = update_configurations(custom_llm)
+
+    progress(0.8, desc="Configuring app..")
+    # List of model selections - easily extendable
+    model_selections = [
+        (openai_model, use_openai),
+        (gemini_model, use_gemini),
+        # Add new models here as ('model_name', use_model),
+    ]
+
+    # Determine the selected LLM provider
+    selected_llm_model = determine_llm_model(model_selections)
+
+    # Update the query engine with the selected LLM model
+    query_engine = AutoQueryEngine.from_defaults(
+        documents=file_documents,
+        use_async=False,
+        system_prompt=custom_llm.instructions,
+        exist_ok=True,
+        overwrite_existing=True,
+        llm_model=selected_llm_model)
+
+    # Complete progress
+    progress(1.0, desc="Completed")  # Complete progress bar
+    create_preview_output = gr.Textbox(
+        """LLM details are updated in configuration tab and LLM App is ready to be previewed 🚀. Start chatting with your custom LLM on the preview 👉"""
+    )
+
+    return create_preview_output, emoji, name, description, instructions
+
+
+def update_configurations(custom_llm):
+    emoji = custom_llm.emoji
+    name = custom_llm.name
+    description = custom_llm.description
+    instructions = custom_llm.instructions
+
+    return gr.Textbox(
+        emoji, interactive=True), gr.Textbox(
+            name, interactive=True), gr.Textbox(
+                description, interactive=True), gr.Textbox(
+                    instructions, interactive=True)
+
+
+def update_app():
+    pass
+
+
+def predict(message, history):
+    chat_response = query_engine.query(message).response
+    return chat_response
+
+
+with gr.Blocks(title="autollm UI", theme=gr.themes.Default(primary_hue=gr.themes.colors.teal)) as demo:
+    gr.Markdown("# LLM Builder")
+    gr.Markdown(
+        """
+        <p style='text-align: center'>
+        Powered by <a href='https://github.com/safevideo/autollm' target='_blank'>autollm</a>
+        </p>
+    """)
+    with gr.Row():
+        with gr.Column():
+            with gr.Tab("Create"):
+                with gr.Accordion(label="LLM Model (default openai gpt-4-0125-preview)", open=False):
+                    with gr.Tab("OpenAI"):
+                        use_openai = gr.Checkbox(value=True, label="Use OpenAI", interactive=True)
+                        openai_model = gr.Dropdown(
+                            label="OpenAI Model",
+                            choices=OPENAI_MODELS,
+                            value=DEFAULT_LLM_MODEL,
+                            interactive=True)
+                        openai_api_key_input = gr.Textbox(label="OPENAI_API_KEY", type="password")
+                    with gr.Tab("Gemini"):
+                        use_gemini = gr.Checkbox(value=False, label="Use Gemini", interactive=True)
+                        gemini_model = gr.Dropdown(
+                            label="Gemini Model", choices=GEMINI_MODELS, value="gemini-pro", interactive=True)
+                        gemini_api_key_input = gr.Textbox(label="GEMINI_API_KEY", type="password")
+                with gr.Accordion(label="Embedding Model API key", open=False):
+                    with gr.Tab("HuggingFace TGI"):
+                        hf_api_key_input = gr.Textbox(label="HF_API_KEY", type="password")
+                what_to_make_area = gr.Textbox(label="What would you like to make?", lines=2)
+
+                with gr.Column(variant="compact"):
+                    with gr.Accordion(label="Add knowledge from files", open=False):
+                        uploaded_files = gr.File(label="Add knowledge from files", file_count="multiple")
+                    with gr.Accordion(label="Add knowledge from folder", open=False):
+                        directory_input = gr.File(
+                            label="Add knowledge from directory", file_count="directory")
+                    with gr.Accordion(label="Add knowledge from webpages", open=False):
+                        webpage_input = gr.Textbox(
+                            lines=2,
+                            info="Enter URLs separated by commas.",
+                            placeholder="https://www.example1.com, https://www.example2.com")
+
+                with gr.Row():
+                    with gr.Column(scale=1, min_width=10):
+                        placeholder = gr.Button(visible=False, interactive=False)
+                    with gr.Column(scale=1, min_width=100):
+                        create_preview_button = gr.Button("Create Preview", variant="primary")
+                create_preview_output = gr.Textbox(
+                    label="Status",
+                    info="Click `Create Preview` 👆 to build preview of the LLM app on the right")
+
+            with gr.Tab("Configure"):
+                with gr.Column(variant="compact"):
+                    detail_html = gr.HTML(
+                        '<a href="https://github.com/safevideo/autollm/blob/main/examples/configs/config.example.yaml">click here for example config</a>'
+                    )
+                    with gr.Accordion(label="Load config file", open=False):
+                        config_file_upload = gr.File(
+                            label="Configurations of LLM, Vector Store..", file_count="single")
+                    emoji = gr.Textbox(label="Emoji")
+                    name = gr.Textbox(label="Name")
+                    description = gr.Textbox(label="Description")
+                    instruction = gr.TextArea(label="Instructions")
+                    with gr.Row():
+                        with gr.Column(scale=1, min_width=10):
+                            placeholder = gr.Button(visible=False, interactive=False)
+                        with gr.Column(scale=1, min_width=100):
+                            update_preview_button = gr.Button("Update Preview", variant="primary")
+                configure_output = gr.Textbox(label="👆 Click `Create Preview` to see preview of the LLM app")
+            with gr.Tab("Export"):
+                # Controls for 'Export' tab
+                hf_api_key = gr.Textbox(label="Hf api key:", type="password")
+                make_db_private = gr.Checkbox(label="Make db private")
+
+        with gr.Column():
+            with gr.Row():
+                download_api_button = gr.Button("Download as API")
+                deploy_button = gr.Button("Deploy to 🤗")
+
+            with gr.Row():
+                with gr.Column():
+                    ai_avatar_image = os.path.join(os.path.dirname(__file__), "serve/avatar.jpg")
+
+                    chatbot = gr.Chatbot(
+                        label="Preview",
+                        bubble_full_width=False,
+                        render=False,
+                        show_copy_button=True,
+                        avatar_images=(None, ai_avatar_image))
+                    chat_interface = gr.ChatInterface(predict, chatbot=chatbot)
+
+        create_preview_button.click(
+            create_app,
+            inputs=[
+                use_openai, openai_model, openai_api_key_input, use_gemini, gemini_model,
+                gemini_api_key_input, what_to_make_area, uploaded_files, webpage_input, config_file_upload
+            ],
+            outputs=[create_preview_output, emoji, name, description, instruction])
+
+        update_preview_button.click(
+            update_app,
+            inputs=[
+                openai_api_key_input, gemini_api_key_input, what_to_make_area, uploaded_files, webpage_input,
+                config_file_upload, emoji, name, description, instruction
+            ],
+            outputs=[configure_output],
+            scroll_to_output=True)
+
+    gr.Markdown(
+        """
+        <p style='text-align: center'>
+        Automatically created by <a href='https://huggingface.co/safevideo' target='_blank'>LLM Builder</a>
+        </p>
+        """)
+
+
+def main():
+    demo.launch()
+
+
+if __name__ == "__main__":
+    main()

autollm/serve/llm_utils.py

@@ -0,0 +1,103 @@
+from typing import Any, Optional
+
+from llama_index.program import LLMTextCompletionProgram
+from pydantic import BaseModel, Field
+
+from autollm import AutoLiteLLM
+
+DEFAULT_LLM_MODEL = "azure/gpt-4-1106"
+DEFAULT_LLM_MAX_TOKENS = 1024
+DEFAULT_LLM_TEMPERATURE = 0.1
+
+
+class CustomLLM(BaseModel):
+    """Data model for custom LLM creation from user prompt."""
+
+    emoji: str = Field(
+        ...,
+        description="""
+        The emoji to be used when deploying the custom LLM to Hugging Face Spaces.
+        """,
+        example="📝",
+    )
+    name: str = Field(
+        ...,
+        description="""
+        The descriptive name of the custom LLM.
+        """,
+        example="Creative Writing Coach",
+    )
+    description: str = Field(
+        ...,
+        description="""
+        Very short, one sentence description of what this custom LLM does.
+        """,
+        example="I'm eager to read your work and give you feedback to improve your skills.",
+    )
+    instructions: str = Field(
+        ...,
+        description="""
+        Very detailed persona instructions for the custom LLM.
+        What does this custom LLM do?
+        How does it behave?
+        What should it avoid doing?
+        How long or short should responses be?
+        """,
+        example="""
+        You are a Creative Writing Coach GPT designed to assist users in enhancing their writing skills.
+        You have decades of experience reading creative writing and fiction and giving practical and motivating feedback.
+        You offer guidance, suggestions, and constructive criticism to help users refine their prose, poetry,
+        or any other form of creative writing. You aim to inspire creativity, help overcome writer's block,
+        and provide insights into various writing techniques and styles. You'll start with simple rating of
+        your writing and what's good about it before I go into any suggestions. Always be positive and encouraging.
+        Ask questions to get more information. Be specific and detailed in your feedback.
+        """,
+    )
+
+
+PROMPT_TEMPLATE_STR = """\
+Your task is to revise the user prompt and create a JSON object \
+in the format of the CustomLLM data model. The JSON object will \
+be used to create a custom LLM model. Ensure the revised prompt \
+maintains the original intent, is clear and detailed, and is \
+adapted to the specific context and task mentioned in the user input.
+
+1. Analyze the basic prompt to understand its primary purpose and context.
+2. Refine the prompt to be clear, detailed, specific, and tailored to the context and task.
+3. Retain the core elements and intent of the original prompt.
+4. Provide an enhanced version of the prompt, ensuring it is optimized for a LLM model interaction.
+
+User prompt: {user_prompt}
+"""
+
+
+def create_custom_llm(user_prompt: str, config: Optional[Any] = None) -> CustomLLM:
+    """Create a custom LLM using the user prompt."""
+    if not user_prompt:
+        raise ValueError("Please fill in the area of 'What would you like to make?'")
+
+    if not config:
+        config = {}
+
+    llm_model = config.get('llm_model', DEFAULT_LLM_MODEL)
+    llm_max_tokens = config.get('llm_max_tokens', DEFAULT_LLM_MAX_TOKENS)
+    llm_temperature = config.get('llm_temperature', DEFAULT_LLM_TEMPERATURE)
+    llm_api_base = config.get('llm_api_base', None)
+
+    llm = AutoLiteLLM.from_defaults(
+        model=llm_model,
+        max_tokens=llm_max_tokens,
+        temperature=llm_temperature,
+        api_base=llm_api_base,
+    )
+
+    program = LLMTextCompletionProgram.from_defaults(
+        output_cls=CustomLLM,
+        prompt_template_str=PROMPT_TEMPLATE_STR,
+        llm=llm,
+        verbose=True,
+    )
+
+    output = program(user_prompt=user_prompt)
+
+    return output

autollm/utils/document_reading.py

@@ -2,6 +2,7 @@
 from pathlib import Path
 from typing import List, Optional, Sequence
 
+from llama_index import download_loader
 from llama_index.readers.file.base import SimpleDirectoryReader
 from llama_index.schema import Document
 
@@ -37,10 +38,13 @@ def read_files_as_documents(
     Returns:
         documents (Sequence[Document]): A sequence of Document objects.
     """
+    JSONReader = download_loader("JSONReader")
+
     # Configure file_extractor to use MarkdownReader for md files
     file_extractor = {
         ".md": MarkdownReader(read_as_single_doc=True),
-        ".pdf": LangchainPDFReader(extract_images=False)
+        ".pdf": LangchainPDFReader(extract_images=False),
+        ".json": JSONReader(),
     }
 
     # Initialize SimpleDirectoryReader

requirements.txt

@@ -2,6 +2,7 @@ llama-index==0.9.27
 litellm==1.16.21
 uvicorn
 fastapi
-python-dotenv
+python-dotenv==1.0.1
 httpx
 lancedb==0.3.4
+gradio==4.14.0

setup.py

@@ -71,4 +71,9 @@ def get_license():
         'Topic :: Scientific/Engineering :: Artificial Intelligence',
         'Topic :: Internet :: WWW/HTTP :: HTTP Servers'
     ],
+    entry_points={
+        'console_scripts': [
+            'autollm=autollm.serve.cli:main',
+        ],
+    },
 )