app.py

import re
import os
import panel as pn
from io import StringIO
from panel.io.mime_render import exec_with_return
from llama_index import (
    VectorStoreIndex,
    SimpleDirectoryReader,
    ServiceContext,
    StorageContext,
    load_index_from_storage,
)
from llama_index.chat_engine import ContextChatEngine
from llama_index.embeddings import OpenAIEmbedding
from llama_index.llms import OpenAI


SYSTEM_PROMPT = (
    "You are a data visualization pro and expert in HoloViz hvplot + holoviews. "
    "Your primary goal is to assist the user in editing based on user requests using best practices. "
    "Simply provide code in code fences (```python). You must have `hvplot_obj` as the last line of code. "
    "Note, data columns are ['sepal_length', 'sepal_width', 'petal_length', 'petal_width', 'species'] and "
    "hvplot is built on top of holoviews--anything you can do with holoviews, you can do "
    "with hvplot. First try to use hvplot **kwargs instead of opts, e.g. `legend='top_right'` "
    "instead of `opts(legend_position='top_right')`. If you need to use opts, you can use "
    "concise version, e.g. `opts(xlabel='Petal Length')` vs `opts(hv.Opts(xlabel='Petal Length'))`"
)

USER_CONTENT_FORMAT = """
Request:
{content}

Code:
```python
{code}
```
""".strip()

DEFAULT_HVPLOT = """
import hvplot.pandas
from bokeh.sampledata.iris import flowers

hvplot_obj = flowers.hvplot(x='petal_length', y='petal_width', by='species', kind='scatter')
hvplot_obj
""".strip()


def exception_handler(exc):
    if retries.value == 0:
        chat_interface.send(f"Can't figure this out: {exc}", respond=False)
        return
    chat_interface.send(f"Fix this error:\n```python\n{exc}\n```")
    retries.value = retries.value - 1


def init_llm(event):
    api_key = event.new
    if not api_key:
        api_key = os.environ.get("OPENAI_API_KEY")
    if not api_key:
        return
    pn.state.cache["llm"] = OpenAI(api_key=api_key)


def create_chat_engine(llm):
    try:
        storage_context = StorageContext.from_defaults(persist_dir="persisted/")
        index = load_index_from_storage(storage_context=storage_context)
    except Exception as exc:
        embed_model = OpenAIEmbedding()
        service_context = ServiceContext.from_defaults(llm=llm, embed_model=embed_model)
        documents = SimpleDirectoryReader(
            input_dir="hvplot_docs", required_exts=[".md"], recursive=True
        ).load_data()
        index = VectorStoreIndex.from_documents(
            documents, service_context=service_context, show_progress=True
        )
        index.storage_context.persist("persisted/")

    retriever = index.as_retriever()
    chat_engine = ContextChatEngine.from_defaults(
        system_prompt=SYSTEM_PROMPT,
        retriever=retriever,
        verbose=True,
    )
    return chat_engine


def callback(content: str, user: str, instance: pn.chat.ChatInterface):
    if "llm" not in pn.state.cache:
        yield "Need to set OpenAI API key first"
        return

    if "engine" not in pn.state.cache:
        engine = pn.state.cache["engine"] = create_chat_engine(pn.state.cache["llm"])
    else:
        engine = pn.state.cache["engine"]

    # new user contents
    user_content = USER_CONTENT_FORMAT.format(
        content=content, code=code_editor.value
    )

    # send user content to chat engine
    agent_response = engine.stream_chat(user_content)

    message = None
    for chunk in agent_response.response_gen:
        message = instance.stream(chunk, message=message, user="OpenAI")

    # extract code
    llm_matches = re.findall(r"```python\n(.*)\n```", message.object, re.DOTALL)
    if llm_matches:
        llm_code = llm_matches[0]
        if llm_code.splitlines()[-1].strip() != "hvplot_obj":
            llm_code += "\nhvplot_obj"
        code_editor.value = llm_code
    retries.value = 2


def update_plot(event):
    with StringIO() as buf:
        hvplot_pane.object = exec_with_return(event.new, stderr=buf)
        buf.seek(0)
        errors = buf.read()
        if errors:
            exception_handler(errors)

pn.extension("codeeditor", sizing_mode="stretch_width", exception_handler=exception_handler)


# instantiate widgets and panes
api_key_input = pn.widgets.PasswordInput(
    placeholder=(
        "Currently subsidized by Andrew, "
        "but you can also pass your own OpenAI API Key"
    )
)
chat_interface = pn.chat.ChatInterface(
    callback=callback,
    show_clear=False,
    show_undo=False,
    show_button_name=False,
    message_params=dict(
        show_reaction_icons=False,
        show_copy_icon=False,
    ),
    height=650,
    callback_exception="verbose",
)
hvplot_pane = pn.pane.HoloViews(
    exec_with_return(DEFAULT_HVPLOT),
    sizing_mode="stretch_both",
)
code_editor = pn.widgets.CodeEditor(
    value=DEFAULT_HVPLOT,
    language="python",
    sizing_mode="stretch_both",
)
retries = pn.widgets.IntInput(value=2, visible=False)
error = pn.widgets.StaticText(visible=False)

# watch for code changes
api_key_input.param.watch(init_llm, "value")
code_editor.param.watch(update_plot, "value")
api_key_input.param.trigger("value")

# lay them out
tabs = pn.Tabs(
    ("Plot", hvplot_pane),
    ("Code", code_editor),
)

sidebar = [api_key_input, chat_interface]
main = [tabs]
template = pn.template.FastListTemplate(
    sidebar=sidebar,
    main=main,
    sidebar_width=600,
    main_layout=None,
    accent_base_color="#fd7000",
    header_background="#fd7000",
    title="Chat with Plot"
)

template.servable()