app.py

import time
import json
import numpy as np
import faiss
import torch
import gradio as gr

from transformers import AutoTokenizer, AutoModel, AutoModelForQuestionAnswering


# -------------------------------------------------------
# CONFIG
# -------------------------------------------------------

# Embedding model for retrieval
EMBED_MODEL = "Desalegnn/Desu-snowflake-arctic-embed-l-v2.0-finetuned-amharic-45k"

# Extractive QA model (generator/reader)
QA_MODEL = "Desalegnn/afroxlmr-amharic-qa"

# Local files in the Space repo (⚠️ make sure names match what you upload)
FAISS_PATH    = "amharic_faiss.bin"    # upload this file
METADATA_PATH = "passage_meta.jsonl"     # upload this file

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print("DEVICE:", DEVICE)


# -------------------------------------------------------
# LOAD MODELS + INDEX + METADATA
# -------------------------------------------------------

# 1) Embedding model
embed_tokenizer = AutoTokenizer.from_pretrained(EMBED_MODEL)
embed_model     = AutoModel.from_pretrained(EMBED_MODEL).to(DEVICE)
embed_model.eval()

# 2) QA model
qa_tokenizer = AutoTokenizer.from_pretrained(QA_MODEL)
qa_model     = AutoModelForQuestionAnswering.from_pretrained(QA_MODEL).to(DEVICE)
qa_model.eval()

# 3) FAISS index
index = faiss.read_index(FAISS_PATH)
print("FAISS dimension:", index.d)

# 4) Passage metadata
metadata = []
with open(METADATA_PATH, "r", encoding="utf-8") as f:
    for line in f:
        line = line.strip()
        if line:
            metadata.append(json.loads(line))

print("Loaded passages:", len(metadata))


# -------------------------------------------------------
# EMBEDDING FUNCTION
# -------------------------------------------------------

@torch.no_grad()
def embed_texts(texts, batch_size=8):
    """
    Embed a list of texts using the Snowflake model (mean-pooled).
    Returns np.ndarray of shape [N, D].
    """
    all_embs = []

    for i in range(0, len(texts), batch_size):
        batch = texts[i:i + batch_size]

        enc = embed_tokenizer(
            batch,
            padding=True,
            truncation=True,
            max_length=256,
            return_tensors="pt",
        ).to(DEVICE)

        out = embed_model(**enc).last_hidden_state  # [B, T, D]
        mask = enc["attention_mask"].unsqueeze(-1)  # [B, T, 1]

        summed = (out * mask).sum(dim=1)           # [B, D]
        counts = mask.sum(dim=1).clamp(min=1e-9)   # [B, 1]
        emb = (summed / counts).cpu().numpy()      # [B, D]

        all_embs.append(emb)

    return np.vstack(all_embs).astype("float32")


# -------------------------------------------------------
# RETRIEVAL
# -------------------------------------------------------

def retrieve_top_k(query, k=5):
    """
    1) Embed query with Snowflake.
    2) Search FAISS index.
    3) Return top-k passages and retrieval latency (ms).
    """
    t0 = time.time()

    query_emb = embed_texts([query])  # [1, D]
    distances, indices = index.search(query_emb, k)

    ret_latency = (time.time() - t0) * 1000.0  # ms

    distances = distances[0]
    indices   = indices[0]

    results = []
    for idx, dist in zip(indices, distances):
        if 0 <= idx < len(metadata):
            meta = metadata[idx]
            results.append(
                {
                    "id": meta.get("id", idx),
                    "text": meta.get("text", ""),
                    "score": float(-dist),  # larger is better
                }
            )

    return results, ret_latency


# -------------------------------------------------------
# EXTRACTIVE QA ON ONE PASSAGE
# -------------------------------------------------------

@torch.no_grad()
def answer_on_context(question, passage):
    """
    Apply AfroXLM-R QA model to (question, passage) and return best span + score.
    """
    enc = qa_tokenizer(
        question,
        passage,
        truncation="only_second",
        max_length=384,
        padding="max_length",
        return_offsets_mapping=True,
        return_tensors="pt",
    )

    input_ids      = enc["input_ids"].to(DEVICE)
    attention_mask = enc["attention_mask"].to(DEVICE)
    offset_mapping = enc["offset_mapping"][0].tolist()
    sequence_ids   = enc.sequence_ids(0)  # 0 = question, 1 = context, None = special

    outputs = qa_model(input_ids=input_ids, attention_mask=attention_mask)

    start_logits = outputs.start_logits[0].cpu().numpy()
    end_logits   = outputs.end_logits[0].cpu().numpy()

    # mask out non-context tokens
    for i, sid in enumerate(sequence_ids):
        if sid != 1:
            start_logits[i] = -1e9
            end_logits[i]   = -1e9

    start_idx = int(np.argmax(start_logits))
    end_idx   = int(np.argmax(end_logits))
    if end_idx < start_idx:
        end_idx = start_idx

    # convert to char positions
    start_char, end_char = offset_mapping[start_idx][0], offset_mapping[end_idx][1]

    if (
        start_char is None
        or end_char is None
        or end_char <= start_char
        or start_char < 0
        or end_char > len(passage)
    ):
        answer_text = ""
    else:
        answer_text = passage[start_char:end_char]

    score = float(start_logits[start_idx] + end_logits[end_idx])

    return answer_text.strip(), score


# -------------------------------------------------------
# RAG PIPELINE: RETRIEVE -> EXTRACTIVE QA
# -------------------------------------------------------

def rag_pipeline(question, k=5):
    """
    1) Retrieve top-k passages.
    2) Run AfroXLM-R QA on each passage.
    3) Select best answer by score.
    4) Return answer, retrieval latency, generator latency, passage snippet.
    """
    # 1) Retrieval
    passages, ret_lat = retrieve_top_k(question, k)

    if not passages:
        return (
            "**Answer:** መረጃ አልተገኘም።",
            f"**Retrieval Latency:** {ret_lat:.2f} ms",
            "**Generator Latency:** 0.00 ms",
            "",
        )

    # 2) QA on each passage
    t0 = time.time()

    best_answer = ""
    best_score  = -1e9
    best_passage_text = ""

    for p in passages:
        ctx = p["text"]
        if not ctx.strip():
            continue

        ans, score = answer_on_context(question, ctx)
        if ans and score > best_score:
            best_score = score
            best_answer = ans
            best_passage_text = ctx

    gen_lat = (time.time() - t0) * 1000.0  # ms

    if not best_answer:
        best_answer = "መልስ አልተገኘም።"

    snippet = best_passage_text[:500] + ("..." if len(best_passage_text) > 500 else "")

    return (
        f"**Answer (AfroXLM-R extractive):** {best_answer}",
        f"**Retrieval Latency:** {ret_lat:.2f} ms",
        f"**Generator Latency (QA):** {gen_lat:.2f} ms",
        snippet,
    )


# -------------------------------------------------------
# GRADIO APP
# -------------------------------------------------------

def gradio_rag(query, k):
    query = (query or "").strip()
    if not query:
        return "Please type a question.", "", "", ""
    return rag_pipeline(query, int(k))


with gr.Blocks() as app:
    gr.Markdown("<h2>🇪🇹 Amharic RAG (Snowflake + AfroXLM-R Extractive QA)</h2>")
    gr.Markdown(
        "Retrieval-Augmented Question Answering: "
        "Snowflake embeddings + FAISS for retrieval, "
        "AfroXLM-R extractive model for answer spans."
    )

    with gr.Row():
        query = gr.Textbox(
            label="Ask an Amharic question",
            lines=2,
            placeholder="ምሳሌ፡ አባይ ወንዝ የት ነው የሚመነጨው?"
        )
        k = gr.Slider(1, 10, value=5, step=1, label="Top-K passages")

    btn = gr.Button("Run RAG")

    out_answer  = gr.Markdown(label="Answer")
    out_retlat  = gr.Markdown(label="Retrieval latency")
    out_genlat  = gr.Markdown(label="Generator latency")
    out_passage = gr.Textbox(label="Retrieved passage snippet", lines=10)

    btn.click(
        gradio_rag,
        inputs=[query, k],
        outputs=[out_answer, out_retlat, out_genlat, out_passage],
    )

app.launch()