app.py

import os
import base64
import pickle
import numpy as np
import gradio as gr
from perplexity import Perplexity
import json

# ─── Configuration ───────────────────────────────────────────────
MODEL_NAME = "pplx-embed-v1-0.6b"
DB_FILE = "case_embeddings.pkl"

# ─── Load Vector Database ────────────────────────────────────────
print(f"Loading vector database from {DB_FILE}...")
if not os.path.exists(DB_FILE):
    raise FileNotFoundError(
        f"{DB_FILE} not found. Please run create_case_db.py first! "
        "(Note: This takes ~12 hours for the full dataset)."
    )

with open(DB_FILE, 'rb') as f:
    database = pickle.load(f)

cases = database["cases"]
legislation = database["legislation"]
mapping = database["mapping"]
corpus_embeddings = database["embeddings"]

# Normalize corpus embeddings for cosine similarity
corpus_norms = np.linalg.norm(corpus_embeddings, axis=1, keepdims=True)
corpus_embeddings_normalized = corpus_embeddings / np.maximum(corpus_norms, 1e-8)

# Pre-calculate indices for each type for faster filtering
case_indices = [i for i, m in enumerate(mapping) if m["type"] == "case"]
leg_indices = [i for i, m in enumerate(mapping) if m["type"] == "legislation"]

case_embeddings = corpus_embeddings_normalized[case_indices]
leg_embeddings = corpus_embeddings_normalized[leg_indices]

print(f"✅ Loaded {len(cases)} clinical cases and {len(legislation)} legislation sections.")


def decode_embedding(b64_string):
    """Decode a base64-encoded int8 embedding to float32."""
    return np.frombuffer(base64.b64decode(b64_string), dtype=np.int8).astype(np.float32)


def format_nursing_reasoning(reasoning):
    """Format the nursing_reasoning dictionary into readable markdown."""
    if not isinstance(reasoning, dict):
        return str(reasoning)
    
    output = ""
    if "Differential Diagnosis" in reasoning:
        output += f"**Differential Diagnosis:**\n{reasoning['Differential Diagnosis']}\n\n"
    if "Diagnostic Tests" in reasoning:
        output += f"**Diagnostic Tests:**\n{reasoning['Diagnostic Tests']}\n\n"
    if "Management Plan" in reasoning:
        output += f"**Management Plan:**\n{reasoning['Management Plan']}\n\n"
    if "Underlying Mechanism" in reasoning:
        output += f"**Underlying Mechanism:**\n{reasoning['Underlying Mechanism']}\n\n"
    if "Relevant Medications" in reasoning:
        output += f"**Relevant Medications:**\n{reasoning['Relevant Medications']}\n\n"
    return output


def perform_search(api_key, query, top_k, search_type):
    """Perform actual semantic search using the Perplexity API."""
    if not api_key or not api_key.strip():
        return "⚠️ **Please enter your Perplexity API Key** above.", ""

    if not query or not query.strip():
        return "Please enter a clinical presentation or legal query.", ""

    try:
        # Embed the query
        client = Perplexity(api_key=api_key.strip())
        response = client.embeddings.create(
            input=[query],
            model=MODEL_NAME
        )
        query_embedding = decode_embedding(response.data[0].embedding)

        # Normalize query
        query_norm = np.linalg.norm(query_embedding)
        query_embedding_normalized = query_embedding / max(query_norm, 1e-8)

        output = f"### 🔍 Results for: *\"{query}\"*\n\n---\n"
        cost_str = ""

        if hasattr(response, 'usage') and response.usage:
            cost_str = f"API Cost: ${response.usage.cost.total_cost:.6f}"

        if search_type == "case":
            # Match against Clinical Cases
            similarities = np.dot(case_embeddings, query_embedding_normalized)
            top_k_indices = np.argsort(similarities)[-top_k:][::-1]

            for i, local_idx in enumerate(top_k_indices):
                score = similarities[local_idx]
                global_idx = case_indices[local_idx]
                case_item = cases[mapping[global_idx]["source_idx"]]
                
                vignette = case_item.get("original_vignette", "")
                reasoning = case_item.get("nursing_reasoning", {})
                case_id = case_item.get("id", "Unknown")

                output += f"#### Result {i+1} — Relevance: {score:.3f} (Case #{case_id})\n"
                output += f"**Patient Presentation:**\n> {vignette}\n\n"
                output += f"<details><summary><b>View Nursing Reasoning</b></summary>\n\n"
                output += f"{format_nursing_reasoning(reasoning)}\n</details>\n\n---\n"

        else:
            # Match against Legislation
            similarities = np.dot(leg_embeddings, query_embedding_normalized)
            top_k_indices = np.argsort(similarities)[-top_k:][::-1]

            for i, local_idx in enumerate(top_k_indices):
                score = similarities[local_idx]
                global_idx = leg_indices[local_idx]
                leg_item = legislation[mapping[global_idx]["source_idx"]]
                
                title = leg_item.get("title", "")
                leg_id = leg_item.get("legislation_id", "")
                text = leg_item.get("text", "")

                output += f"#### Result {i+1} — Relevance: {score:.3f} (Act: {leg_id})\n"
                output += f"⚖️ **{title}**\n\n"
                output += f"{text}\n\n---\n"

        return output, cost_str

    except Exception as e:
        error_msg = str(e)
        if "401" in error_msg or "auth" in error_msg.lower():
            return "❌ **Invalid API Key.** Check your key and try again.", ""
        return f"❌ Error: {error_msg}", ""


def search_cases(api_key, query, top_k):
    return perform_search(api_key, query, top_k, search_type="case")

def search_legislation(api_key, query, top_k):
    return perform_search(api_key, query, top_k, search_type="legislation")


# ─── Gradio UI ───────────────────────────────────────────────────
with gr.Blocks(title="NurseLex-Match") as app:
    gr.Markdown(
        """
        # 🩺 NurseLex-Match
        ### Clinical Case Similarity & Legal Lookup
        *Powered by Perplexity Embeddings (`pplx-embed-v1-0.6b`)*
        
        Retrieve similar historical cases from the **NurseReason-Dataset** or search UK nursing law in **NurseLex**.
        """
    )

    with gr.Accordion("🔑 API Key (BYOK — Bring Your Own Key)", open=True):
        gr.Markdown(
            "Your key is **never stored** — it is used only for this session to query the embeddings API. "
        )
        api_key_input = gr.Textbox(
            label="Perplexity API Key",
            placeholder="pplx-...",
            type="password",
            lines=1
        )
        api_cost = gr.Markdown("*API Cost: $0.000000*")

    with gr.Tabs():
        # TAB 1: Clinical Cases
        with gr.TabItem("🏥 Clinical Case Matcher"):
            with gr.Row():
                with gr.Column(scale=3):
                    case_search_input = gr.Textbox(
                        label="Describe the patient presentation:",
                        placeholder="e.g., 72-year-old presenting with acute confusion and suspected UTI...",
                        lines=3
                    )
                with gr.Column(scale=1):
                    case_top_k = gr.Slider(minimum=1, maximum=10, step=1, value=3, label="Results")
                    case_search_btn = gr.Button("🔍 Find Similar Cases", variant="primary")
            
            gr.Examples(
                examples=[
                    "Patient is a 45-year-old female complaining of sharp left lower quadrant abdominal pain radiating to the back.",
                    "Elderly patient with a history of heart failure presenting with increased shortness of breath and pitting edema.",
                    "Post-operative patient day 2 showing signs of infection at the wound site with low-grade fever."
                ],
                inputs=case_search_input
            )
            case_results = gr.Markdown("*Similar cases will appear here...*")

        # TAB 2: Nursing Legislation
        with gr.TabItem("⚖️ Legal Lookup"):
            with gr.Row():
                with gr.Column(scale=3):
                    leg_search_input = gr.Textbox(
                        label="Describe the legal context or policy question:",
                        placeholder="e.g., What are the rules regarding compulsory admission under the Mental Health Act?",
                        lines=3
                    )
                with gr.Column(scale=1):
                    leg_top_k = gr.Slider(minimum=1, maximum=10, step=1, value=3, label="Results")
                    leg_search_btn = gr.Button("🔍 Search Legislation", variant="primary")
            
            gr.Examples(
                examples=[
                    "When can a nurse legally refuse to participate in a procedure?",
                    "What is the required staffing ratio for an intensive care unit?",
                    "Regulations regarding the administration of controlled drugs."
                ],
                inputs=leg_search_input
            )
            leg_results = gr.Markdown("*Relevant legislation will appear here...*")

    # Wire up search events
    case_search_btn.click(
        fn=search_cases,
        inputs=[api_key_input, case_search_input, case_top_k],
        outputs=[case_results, api_cost]
    )
    case_search_input.submit(
        fn=search_cases,
        inputs=[api_key_input, case_search_input, case_top_k],
        outputs=[case_results, api_cost]
    )

    leg_search_btn.click(
        fn=search_legislation,
        inputs=[api_key_input, leg_search_input, leg_top_k],
        outputs=[leg_results, api_cost]
    )
    leg_search_input.submit(
        fn=search_legislation,
        inputs=[api_key_input, leg_search_input, leg_top_k],
        outputs=[leg_results, api_cost]
    )

if __name__ == "__main__":
    print("Starting NurseLex-Match...")
    app.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)