app.py

# app.py – Unified Panel App with Semantic Search + Filterable Tabulator

import os, io, gc
import panel as pn
import pandas as pd
import boto3, torch
import psycopg2
from sentence_transformers import SentenceTransformer, util

pn.extension('tabulator')

# ──────────────────────────────────────────────────────────────────────
# 1) Database and Resource Loading
# ──────────────────────────────────────────────────────────────────────
DB_HOST = os.getenv("DB_HOST")
DB_PORT = os.getenv("DB_PORT", "5432")
DB_NAME = os.getenv("DB_NAME")
DB_USER = os.getenv("DB_USER")
DB_PASSWORD = os.getenv("DB_PASSWORD")

@pn.cache()
def get_data():
    conn = psycopg2.connect(
        host=DB_HOST, port=DB_PORT,
        dbname=DB_NAME, user=DB_USER, password=DB_PASSWORD,
        sslmode="require"
    )
    df_ = pd.read_sql_query("""
        SELECT id, country, year, section,
               question_code, question_text,
               answer_code,  answer_text
          FROM survey_info;
    """, conn)
    conn.close()
    return df_

df = get_data()

@pn.cache()
def load_embeddings():
    BUCKET, KEY = "cgd-embeddings-bucket", "survey_info_embeddings.pt"
    buf = io.BytesIO()
    boto3.client("s3").download_fileobj(BUCKET, KEY, buf)
    buf.seek(0)
    ckpt = torch.load(buf, map_location="cpu")
    buf.close(); gc.collect()
    return ckpt["ids"], ckpt["embeddings"]

@pn.cache()
def get_st_model():
    return SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2", device="cpu")

@pn.cache()
def get_semantic_resources():
    model = get_st_model()
    ids_list, emb_tensor = load_embeddings()
    return model, ids_list, emb_tensor

# ──────────────────────────────────────────────────────────────────────
# 2) Widgets
# ──────────────────────────────────────────────────────────────────────
country_opts = sorted(df["country"].dropna().unique())
year_opts = sorted(df["year"].dropna().unique())

w_countries = pn.widgets.MultiSelect(name="Countries", options=country_opts)
w_years = pn.widgets.MultiSelect(name="Years", options=year_opts)
w_keyword = pn.widgets.TextInput(name="Keyword Search", placeholder="Search questions or answers")
w_group = pn.widgets.Checkbox(name="Group by Question Text", value=False)

w_semquery = pn.widgets.TextInput(name="Semantic Query")
w_search_button = pn.widgets.Button(name="Semantic Search", button_type="primary")

# ──────────────────────────────────────────────────────────────────────
# 3) Unified Results Table (Tabulator)
# ──────────────────────────────────────────────────────────────────────
result_table = pn.widgets.Tabulator(
    pagination='remote',
    page_size=15,
    sizing_mode="stretch_width",
    layout='fit_columns',
    show_index=False,
)

@pn.depends(w_countries, w_years, w_keyword, w_group, watch=True)
def update_table(countries, years, keyword, group):
    filt = df.copy()
    if countries:
        filt = filt[filt["country"].isin(countries)]
    if years:
        filt = filt[filt["year"].isin(years)]
    if keyword:
        filt = filt[
            filt["question_text"].str.contains(keyword, case=False, na=False) |
            filt["answer_text"].str.contains(keyword, case=False, na=False) |
            filt["question_code"].astype(str).str.contains(keyword, case=False, na=False)
        ]

    if group:
        grouped = (
            filt.groupby("question_text")
            .agg({
                "country": lambda x: sorted(set(x)),
                "year": lambda x: sorted(set(x)),
                "answer_text": lambda x: list(x)[:3]
            })
            .reset_index()
            .rename(columns={
                "country": "Countries",
                "year": "Years",
                "answer_text": "Sample Answers"
            })
        )
        result_table.value = grouped
    else:
        result_table.value = filt[["country", "year", "question_text", "answer_text"]]

def semantic_search(event=None):
    query = w_semquery.value.strip()
    if not query:
        return

        # Step 1: Filter the full dataframe
    filt = df.copy()
    if w_countries.value:
        filt = filt[filt["country"].isin(w_countries.value)]
    if w_years.value:
        filt = filt[filt["year"].isin(w_years.value)]
    if w_keyword.value:
        filt = filt[
            filt["question_text"].str.contains(w_keyword.value, case=False, na=False) |
            filt["answer_text"].str.contains(w_keyword.value, case=False, na=False) |
            filt["question_code"].astype(str).str.contains(w_keyword.value, case=False, na=False)
        ]

    # Step 2: Load only embeddings for the filtered rows
    model, ids_list, emb_tensor = get_semantic_resources()
    
    # Create a mask for filtered IDs
    filtered_ids = filt["id"].tolist()
    id_to_index = {id_: i for i, id_ in enumerate(ids_list)}
    filtered_indices = [id_to_index[id_] for id_ in filtered_ids if id_ in id_to_index]
    
    # Subset the embedding tensor
    filtered_embs = emb_tensor[filtered_indices]
    
    # Step 3: Semantic search only within filtered subset
    q_vec = model.encode(query, convert_to_tensor=True, device="cpu").cpu()
    sims = util.cos_sim(q_vec, filtered_embs)[0]
    top_vals, top_idx = torch.topk(sims, k=50)
    
    top_filtered_ids = [filtered_ids[i] for i in top_idx.tolist()]
    sem_rows = filt[filt["id"].isin(top_filtered_ids)].copy()
    score_map = dict(zip(top_filtered_ids, top_vals.tolist()))
    sem_rows["Score"] = sem_rows["id"].map(score_map)
    sem_rows = sem_rows.sort_values("Score", ascending=False)
    
    # Final output
    result_table.value = sem_rows[["Score", "country", "year", "question_text", "answer_text"]]


    filt = df.copy()
    if w_countries.value:
        filt = filt[filt["country"].isin(w_countries.value)]
    if w_years.value:
        filt = filt[filt["year"].isin(w_years.value)]
    if w_keyword.value:
        filt = filt[
            filt["question_text"].str.contains(w_keyword.value, case=False, na=False) |
            filt["answer_text"].str.contains(w_keyword.value, case=False, na=False) |
            filt["question_code"].astype(str).str.contains(w_keyword.value, case=False, na=False)
        ]

    remainder = filt.loc[~filt["id"].isin(sem_ids)].copy()
    remainder["Score"] = ""

    combined = pd.concat([sem_rows, remainder], ignore_index=True)
    result_table.value = combined[["Score", "country", "year", "question_text", "answer_text"]]

w_search_button.on_click(semantic_search)

# ──────────────────────────────────────────────────────────────────────
# 4) Layout
# ──────────────────────────────────────────────────────────────────────
sidebar = pn.Column(
    "## 🔎 Filters",
    w_countries, w_years, w_keyword, w_group,
    pn.Spacer(height=20),
    "## 🧠 Semantic Search",
    w_semquery, w_search_button,
    width=300
)

main = pn.Column(
    pn.pane.Markdown("## 🌍 CGD Survey Explorer"),
    result_table
)

pn.template.FastListTemplate(
    title="CGD Survey Explorer",
    sidebar=sidebar,
    main=main,
    theme_toggle=True,
).servable()