app.py

"""
app.py — FECB RAG Search Application

Loads a pre-built FAISS index (produced by ingest.py) and provides a
Gradio interface for semantic search and AI-assisted Q&A over your PDF documents.

Environment variables:
    ANTHROPIC_API_KEY  — Anthropic API key (required)
    CLAUDE_MODEL       — Claude model ID (default: claude-sonnet-4-6)
    EMBED_MODEL        — Embedding model (default: BAAI/bge-small-en-v1.5)
    TOP_K              — Max documents to retrieve (default: 5)
    INDEX_DIR          — Path to FAISS index (default: faiss_index)
    META_FILE          — Path to metadata JSON (default: metadata.json)

Run:
    python app.py
"""

import json
import os
import re
from pathlib import Path

import anthropic
import gradio as gr
from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings

# ── Config ────────────────────────────────────────────────────────────────────
EMBED_MODEL = os.getenv("EMBED_MODEL", "BAAI/bge-small-en-v1.5")
CLAUDE_MODEL = os.getenv("CLAUDE_MODEL", "claude-sonnet-4-6")
API_KEY     = os.getenv("ANTHROPIC_API_KEY")
TOP_K       = int(os.getenv("TOP_K", "8"))
INDEX_DIR   = Path(os.getenv("INDEX_DIR", "faiss_index"))
META_FILE   = Path(os.getenv("META_FILE", "metadata.json"))

SYSTEM_PROMPT = (
    "You are a knowledgeable research assistant. You help users find relevant "
    "information from a document collection and synthesize key findings. "
    "When answering, cite the specific document(s) by their bracketed number [N]. "
    "Be concise and precise. If the context doesn't contain enough information "
    "to answer fully, say so clearly."
)

# ── Load resources ────────────────────────────────────────────────────────────
print(f"Loading embedding model: {EMBED_MODEL}")
_embeddings = HuggingFaceEmbeddings(
    model_name=EMBED_MODEL,
    model_kwargs={"device": "cpu"},
    encode_kwargs={"normalize_embeddings": True},
)

if not INDEX_DIR.exists():
    raise FileNotFoundError(
        f"FAISS index not found at '{INDEX_DIR}'. "
        "Run 'python ingest.py' first to build the index from your PDFs."
    )

print(f"Loading FAISS index from: {INDEX_DIR}")
_vectorstore = FAISS.load_local(
    str(INDEX_DIR), _embeddings, allow_dangerous_deserialization=True
)

_metadata: dict[str, dict] = {}
if META_FILE.exists():
    print(f"Loading metadata from: {META_FILE}")
    with open(META_FILE, encoding="utf-8") as f:
        for record in json.load(f):
            _metadata[record["doc_id"]] = record
    print(f"  Loaded metadata for {len(_metadata)} documents")
else:
    print(f"  [WARN] {META_FILE} not found — document names will be inferred from IDs")

if not API_KEY:
    raise EnvironmentError("ANTHROPIC_API_KEY is not set. Export it before running.")

_client = anthropic.Anthropic(api_key=API_KEY)
print(f"Claude model: {CLAUDE_MODEL}")
print("Ready.\n")


# ── RAG helpers ───────────────────────────────────────────────────────────────

def retrieve(query: str, n_chunks: int) -> list[tuple]:
    """Return the top n_chunks most relevant chunks, allowing multiple per document."""
    raw = _vectorstore.similarity_search_with_score(query, k=n_chunks)
    return sorted(raw, key=lambda x: x[1])


def build_context(hits: list[tuple]) -> str:
    parts = []
    for i, (doc, _) in enumerate(hits, 1):
        doc_id   = doc.metadata.get("doc_id", f"doc_{i}")
        filename = doc.metadata.get("filename", f"{doc_id}.pdf")
        excerpt  = doc.page_content.strip()
        parts.append(f"[{i}] {filename} (ID: {doc_id})\n{excerpt}")
    return "\n\n---\n\n".join(parts)


def ask_claude(query: str, context: str) -> str:
    user_content = (
        f"Using the document excerpts below, answer the following question. "
        f"Cite documents by their bracketed number.\n\n"
        f"Question: {query}\n\nContext:\n{context}"
    )
    try:
        message = _client.messages.create(
            model=CLAUDE_MODEL,
            max_tokens=800,
            system=SYSTEM_PROMPT,
            messages=[{"role": "user", "content": user_content}],
        )
        return message.content[0].text.strip()
    except anthropic.APIError as exc:
        return (
            f"Could not reach Claude ({exc}).\n\n"
            "Check that **ANTHROPIC_API_KEY** is set and valid."
        )


def cosine_to_pct(score: float) -> str:
    """Convert FAISS L2 distance (normalised embeddings) to 0–100% relevance."""
    pct = (1.0 - min(max(score, 0.0), 2.0) / 2.0) * 100
    return f"{pct:.1f}%"


# ── Main search function ──────────────────────────────────────────────────────

def rag_search(query: str, n_docs: int) -> tuple[str, str]:
    query = query.strip()
    if not query:
        return "Please enter a question or keyword.", ""

    hits = retrieve(query, n_docs)
    if not hits:
        return "No relevant documents found. Try different keywords.", ""

    context = build_context(hits)
    answer  = ask_claude(query, context)

    cards = []
    for i, (doc, score) in enumerate(hits, 1):
        doc_id   = doc.metadata.get("doc_id", f"doc_{i}")
        filename = doc.metadata.get("filename", f"{doc_id}.pdf")
        rel      = cosine_to_pct(score)
        snippet  = doc.page_content.replace("\n", " ").strip()[:350]

        cards.append(
            f"### [{i}] {filename}\n"
            f"**Relevance:** {rel}  \n"
            f"**ID:** {doc_id}  \n"
            f"> {snippet}…"
        )

    return answer, "\n\n---\n\n".join(cards)


# ── Gradio UI ─────────────────────────────────────────────────────────────────

with gr.Blocks(title="FECB Document Search") as demo:

    gr.Markdown(
        """
# FECB Document Search — AI-Powered RAG

Search your document collection using semantic AI search.
Ask a question or enter keywords; the app retrieves the most relevant
documents and generates a synthesised answer with citations.

> **Powered by** `BAAI/bge-small-en-v1.5` embeddings · Claude via Anthropic API
        """
    )

    with gr.Row():
        with gr.Column(scale=5):
            query_box = gr.Textbox(
                label="Question or keywords",
                placeholder="e.g. 'What are the main findings about X?'",
                lines=2,
                elem_id="query-box",
            )
        with gr.Column(scale=1, min_width=160):
            n_slider = gr.Slider(
                minimum=3, maximum=20, value=TOP_K, step=1,
                label="Chunks to retrieve",
            )

    search_btn = gr.Button("Search", variant="primary", size="lg")

    gr.Markdown("---")

    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown("### AI Answer")
            answer_md = gr.Markdown(value="*Results will appear here after searching.*")

        with gr.Column(scale=3):
            gr.Markdown("### Relevant Documents")
            papers_md = gr.Markdown(value="")

    search_btn.click(rag_search, [query_box, n_slider], [answer_md, papers_md])
    query_box.submit(rag_search, [query_box, n_slider], [answer_md, papers_md])


if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,
        theme=gr.themes.Soft(primary_hue="blue", font=gr.themes.GoogleFont("Inter")),
        css="""
            .gradio-container { max-width: 1100px; margin: auto; }
            #query-box textarea { font-size: 16px; }
        """,
    )