ui.py

import base64
import streamlit as st
from datetime import datetime
from scrape import scrape_multiple
from search import get_search_results
from llm_utils import BufferedStreamingHandler, get_model_choices
from llm import get_llm, refine_query, filter_results, generate_summary


def _render_pipeline_error(stage: str, err: Exception) -> None:
    message = str(err).strip() or err.__class__.__name__
    lower_msg = message.lower()
    hints = [
        "- Confirm the relevant API key is set in your `.env` or shell before launching Streamlit.",
        "- Keys copied from dashboards often include hidden spaces; re-copy if authentication keeps failing.",
        "- Restart the app after updating environment variables so the new values are picked up.",
    ]

    if any(token in lower_msg for token in ("anthropic", "x-api-key", "invalid api key", "authentication")):
        hints.insert(0, "- Claude/Anthropic models require a valid `ANTHROPIC_API_KEY`.")
    elif "openrouter" in lower_msg:
        hints.insert(0, "- OpenRouter models require `OPENROUTER_API_KEY` and a reachable OpenRouter endpoint.")
    elif "openai" in lower_msg or "gpt" in lower_msg:
        hints.insert(0, "- OpenAI models require `OPENAI_API_KEY`, or configure `OPENAI_BASE_URL`/`OPENAI_URL` for a compatible endpoint.")
    elif "google" in lower_msg or "gemini" in lower_msg:
        hints.insert(0, "- Google Gemini models need `GOOGLE_API_KEY` or Application Default Credentials.")

    st.error(
        "❌ Failed to {}.\n\nError: {}\n\n{}".format(
            stage,
            message,
            "\n".join(hints),
        )
    )
    st.stop()


# Cache expensive backend calls
@st.cache_data(ttl=200, show_spinner=False)
def cached_search_results(refined_query: str, threads: int):
    return get_search_results(refined_query.replace(" ", "+"), max_workers=threads)


@st.cache_data(ttl=200, show_spinner=False)
def cached_scrape_multiple(filtered: list, threads: int):
    return scrape_multiple(filtered, max_workers=threads)


# Streamlit page configuration
st.set_page_config(
    page_title="Robin: AI-Powered Dark Web OSINT Tool",
    page_icon="🕵️‍♂️",
    initial_sidebar_state="expanded",
)

# Custom CSS for styling
st.markdown(
    """
    <style>
            .colHeight {
                max-height: 40vh;
                overflow-y: auto;
                text-align: center;
            }
            .pTitle {
                font-weight: bold;
                color: #FF4B4B;
                margin-bottom: 0.5em;
            }
            .aStyle {
                font-size: 18px;
                font-weight: bold;
                padding: 5px;
                padding-left: 0px;
                text-align: center;
            }
    </style>""",
    unsafe_allow_html=True,
)


# Sidebar
st.sidebar.title("Robin")
st.sidebar.text("AI-Powered Dark Web OSINT Tool")
st.sidebar.markdown(
    """Made by [Apurv Singh Gautam](https://www.linkedin.com/in/apurvsinghgautam/)"""
)
st.sidebar.subheader("Settings")
model_options = get_model_choices()
default_model_index = (
    next(
        (idx for idx, name in enumerate(model_options) if name.lower() in {"gpt-5-mini", "gpt-4.1"}),
        0,
    )
    if model_options
    else 0
)
selected_model = None
if model_options:
    selected_model = st.sidebar.selectbox(
        "Select LLM Model",
        model_options,
        index=default_model_index,
        key="model_select",
    )
else:
    st.sidebar.info("No detected models. Enter a custom model name below.")

custom_model = st.sidebar.text_input(
    "Custom LLM Model (optional)",
    placeholder="e.g. gpt-4.1 or openai/gpt-oss-120b",
    key="custom_model_input",
).strip()
model = custom_model or selected_model
if any(name not in {"gpt4o", "gpt-4.1", "claude-3-5-sonnet-latest", "llama3.1", "gemini-2.5-flash"} for name in model_options):
    st.sidebar.caption("Locally detected Ollama models are automatically added to this list.")
threads = st.sidebar.slider("Scraping Threads", 1, 16, 4, key="thread_slider")


# Main UI - title and input
st.markdown("## Robin: AI-Powered Dark Web OSINT Tool")

# Display text box and button
with st.form("search_form", clear_on_submit=True):
    col_input, col_button = st.columns([10, 1])
    query = col_input.text_input(
        "Enter Dark Web Search Query",
        placeholder="Enter Dark Web Search Query",
        label_visibility="collapsed",
        key="query_input",
    )
    run_button = col_button.form_submit_button("Run")

# Display a status message
status_slot = st.empty()
# Pre-allocate three placeholders-one per card
cols = st.columns(3)
p1, p2, p3 = [col.empty() for col in cols]
# Summary placeholders
summary_container_placeholder = st.empty()


# Process the query
if run_button and query:
    if not model:
        st.error("Please select a model or enter a custom model name.")
        st.stop()

    # clear old state
    for k in ["refined", "results", "filtered", "scraped", "streamed_summary"]:
        st.session_state.pop(k, None)

    # Stage 1 - Load LLM
    with status_slot.container():
        with st.spinner("🔄 Loading LLM..."):
            try:
                llm = get_llm(model)
            except Exception as e:
                _render_pipeline_error("load the selected LLM", e)

    # Stage 2 - Refine query
    with status_slot.container():
        with st.spinner("🔄 Refining query..."):
            try:
                st.session_state.refined = refine_query(llm, query)
            except Exception as e:
                _render_pipeline_error("refine the query", e)
    p1.container(border=True).markdown(
        f"<div class='colHeight'><p class='pTitle'>Refined Query</p><p>{st.session_state.refined}</p></div>",
        unsafe_allow_html=True,
    )

    # Stage 3 - Search dark web
    with status_slot.container():
        with st.spinner("🔍 Searching dark web..."):
            st.session_state.results = cached_search_results(
                st.session_state.refined, threads
            )
    p2.container(border=True).markdown(
        f"<div class='colHeight'><p class='pTitle'>Search Results</p><p>{len(st.session_state.results)}</p></div>",
        unsafe_allow_html=True,
    )

    # Stage 4 - Filter results
    with status_slot.container():
        with st.spinner("🗂️ Filtering results..."):
            st.session_state.filtered = filter_results(
                llm, st.session_state.refined, st.session_state.results
            )
    p3.container(border=True).markdown(
        f"<div class='colHeight'><p class='pTitle'>Filtered Results</p><p>{len(st.session_state.filtered)}</p></div>",
        unsafe_allow_html=True,
    )

    # Stage 5 - Scrape content
    with status_slot.container():
        with st.spinner("📜 Scraping content..."):
            st.session_state.scraped = cached_scrape_multiple(
                st.session_state.filtered, threads
            )

    # Stage 6 - Summarize
    # 6a) Prepare session state for streaming text
    st.session_state.streamed_summary = ""

    # 6c) UI callback for each chunk
    def ui_emit(chunk: str):
        st.session_state.streamed_summary += chunk
        summary_slot.markdown(st.session_state.streamed_summary)

    with summary_container_placeholder.container():  # border=True, height=450):
        hdr_col, btn_col = st.columns([4, 1], vertical_alignment="center")
        with hdr_col:
            st.subheader(":red[Investigation Summary]", anchor=None, divider="gray")
        summary_slot = st.empty()

    # 6d) Inject your two callbacks and invoke exactly as before
    with status_slot.container():
        with st.spinner("✍️ Generating summary..."):
            stream_handler = BufferedStreamingHandler(ui_callback=ui_emit)
            llm.callbacks = [stream_handler]
            _ = generate_summary(llm, query, st.session_state.scraped)

    with btn_col:
        now = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
        fname = f"summary_{now}.md"
        b64 = base64.b64encode(st.session_state.streamed_summary.encode()).decode()
        href = f'<div class="aStyle">📥 <a href="data:file/markdown;base64,{b64}" download="{fname}">Download</a></div>'
        st.markdown(href, unsafe_allow_html=True)
    status_slot.success("✔️ Pipeline completed successfully!")