Upload 3 files

app_urdu.py

@@ -0,0 +1,209 @@
+import os, json, re, time
+import numpy as np
+import gradio as gr
+from sentence_transformers import SentenceTransformer, CrossEncoder
+from rank_bm25 import BM25Okapi
+from pinecone import Pinecone
+from huggingface_hub import InferenceClient
+
+# CONFIG
+PINECONE_API_KEY = os.getenv("PINECONE_API_KEY")
+HF_TOKEN         = os.getenv("HF_TOKEN")
+PINECONE_INDEX   = "rag-nlp-project"
+
+# LOAD RESOURCES
+print("Loading resources...")
+
+with open("chunks_urdu_recursive.json", encoding="utf-8") as f:
+    ALL_CHUNKS = json.load(f)
+
+tokenized = [c["text"].lower().split() for c in ALL_CHUNKS]
+bm25 = BM25Okapi(tokenized)
+
+embedder = SentenceTransformer("paraphrase-multilingual-MiniLM-L12-v2")
+reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")
+
+pc = Pinecone(api_key=PINECONE_API_KEY)
+pine_index = pc.Index(PINECONE_INDEX)
+
+llm = InferenceClient(token=HF_TOKEN)
+print("All ready!")
+
+# RETRIEVAL
+def semantic_search(query, top_k=20):
+    qvec = embedder.encode(query).tolist()
+    res = pine_index.query(vector=qvec, top_k=top_k,
+                           include_metadata=True, namespace="urdu_recursive")
+    return [{"id": m["id"], "text": m["metadata"]["text"],
+             "title": m["metadata"]["title"], "score": m["score"]}
+            for m in res["matches"]]
+
+def bm25_search(query, top_k=20):
+    tokens = query.lower().split()
+    scores = bm25.get_scores(tokens)
+    top_idx = np.argsort(scores)[::-1][:top_k]
+    return [{"id": ALL_CHUNKS[i]["id"], "text": ALL_CHUNKS[i]["text"],
+             "title": ALL_CHUNKS[i]["title"], "score": float(scores[i])}
+            for i in top_idx if scores[i] > 0]
+
+def rrf_fuse(lists_of_results, k=60):
+    scores, data = {}, {}
+    for results in lists_of_results:
+        for rank, item in enumerate(results):
+            did = item["id"]
+            scores[did] = scores.get(did, 0) + 1.0 / (k + rank + 1)
+            data[did] = {"text": item["text"], "title": item["title"]}
+    ranked = sorted(scores, key=lambda x: scores[x], reverse=True)
+    return [{"id": d, "rrf_score": scores[d], **data[d]} for d in ranked]
+
+def cross_encoder_rerank(query, candidates, top_k=5):
+    if not candidates:
+        return []
+    pool = candidates[:30]
+    pairs = [(query, c["text"]) for c in pool]
+    ce_scores = reranker.predict(pairs)
+    for i, s in enumerate(ce_scores):
+        pool[i]["ce_score"] = float(s)
+    pool.sort(key=lambda x: x["ce_score"], reverse=True)
+    return pool[:top_k]
+
+# LLM
+def call_llm(prompt, max_tokens=512, temperature=0.3):
+    for model in ["mistralai/Mistral-7B-Instruct-v0.2", "meta-llama/Meta-Llama-3-8B-Instruct"]:
+        try:
+            resp = llm.chat_completion(
+                model=model,
+                messages=[{"role": "user", "content": prompt}],
+                max_tokens=max_tokens, temperature=temperature
+            )
+            return resp.choices[0].message.content.strip()
+        except:
+            continue
+    return "[LLM Error]"
+
+def generate_answer(query, contexts):
+    ctx = "\n\n".join([f"{i+1}. [{c['title']}] {c['text']}" for i, c in enumerate(contexts)])
+    prompt = f"""درج ذیل معلومات کی بنیاد پر:
+
+{ctx}
+
+سوال کا تفصیلی جواب دیں: {query}
+اگر معلومات سوال سے متعلق نہیں ہیں تو بتائیں کہ ڈیٹابیس میں مناسب جواب نہیں ملا۔"""
+    return call_llm(prompt)
+
+# EVALUATION
+def eval_faithfulness(answer, contexts):
+    context_str = "\n".join([c["text"] for c in contexts])[:3000]
+    claims_raw = call_llm(
+        f"Extract all factual claims as a numbered list.\n\nAnswer: {answer}\n\nClaims:",
+        max_tokens=400, temperature=0.1
+    )
+    claims = [re.sub(r"^[\d]+[\.\)]\s*", "", l.strip())
+              for l in claims_raw.split("\n")
+              if len(re.sub(r"^[\d]+[\.\)]\s*", "", l.strip())) > 15]
+    if not claims:
+        return 1.0, "No claims extracted."
+    supported = 0
+    details = []
+    for claim in claims[:8]:
+        verdict = call_llm(
+            f"Is this claim supported by the context? Reply ONLY 'SUPPORTED' or 'NOT SUPPORTED'.\n\n"
+            f"Context: {context_str}\n\nClaim: {claim}\n\nVerdict:",
+            max_tokens=10, temperature=0.1
+        ).upper()
+        ok = "SUPPORTED" in verdict and "NOT" not in verdict
+        if ok:
+            supported += 1
+        details.append(f"{'[Y]' if ok else '[N]'} {claim}")
+    return supported / len(claims[:8]), "\n".join(details)
+
+def eval_relevancy(query, answer):
+    qs_raw = call_llm(
+        f"Generate exactly 3 questions that this answer directly addresses. "
+        f"One per line, no numbering.\n\nAnswer: {answer}\n\nQuestions:",
+        max_tokens=200, temperature=0.3
+    )
+    questions = [re.sub(r"^[\d]+[\.\)]\s*", "", l.strip())
+                 for l in qs_raw.split("\n")
+                 if len(re.sub(r"^[\d]+[\.\)]\s*", "", l.strip())) > 10][:3]
+    if not questions:
+        return 0.0, "Could not generate questions."
+    embs = embedder.encode([query] + questions)
+    q_emb = embs[0]
+    sims, detail_lines = [], []
+    for i, q in enumerate(questions):
+        sim = float(np.dot(q_emb, embs[i+1]) /
+                    (np.linalg.norm(q_emb) * np.linalg.norm(embs[i+1])))
+        sims.append(sim)
+        detail_lines.append(f"  Q{i+1}: {q}  (sim={sim:.3f})")
+    return float(np.mean(sims)), "\n".join(detail_lines)
+
+# MAIN PIPELINE
+def run_query(query, run_eval):
+    if not query.strip():
+        return "براہ کرم سوال درج کریں۔", "", "", ""
+
+    t0 = time.time()
+    sem = semantic_search(query)
+    kw = bm25_search(query)
+    fused = rrf_fuse([sem, kw])
+    reranked = cross_encoder_rerank(query, fused)
+    t_retrieve = time.time() - t0
+
+    t1 = time.time()
+    answer = generate_answer(query, reranked)
+    t_generate = time.time() - t1
+
+    ctx_display = ""
+    for i, c in enumerate(reranked):
+        ctx_display += f"**[{i+1}] {c['title']}**  (score: {c.get('ce_score', 0):.3f})\n"
+        ctx_display += f"{c['text']}\n\n---\n\n"
+
+    scores_display = ""
+    t_eval = 0
+    if run_eval:
+        t2 = time.time()
+        faith_score, faith_detail = eval_faithfulness(answer, reranked)
+        rel_score, rel_detail = eval_relevancy(query, answer)
+        t_eval = time.time() - t2
+        scores_display = (
+            f"### Faithfulness: {faith_score:.0%}\n{faith_detail}\n\n"
+            f"### Relevancy: {rel_score:.0%}\n{rel_detail}"
+        )
+    else:
+        scores_display = "*(Check the box to run evaluation)*"
+
+    timing = (f"Retrieval: {t_retrieve:.2f}s | Generation: {t_generate:.2f}s | "
+              f"Evaluation: {t_eval:.2f}s | Total: {t_retrieve + t_generate + t_eval:.2f}s")
+    return answer, ctx_display, scores_display, timing
+
+# GRADIO UI
+with gr.Blocks(title="RAG Q&A — Urdu") as demo:
+    gr.Markdown(
+        "# اردو سوال و جواب کا نظام\n"
+        "*Hybrid Search (BM25 + Semantic + RRF) - Cross-Encoder Reranking - LLM-as-a-Judge*"
+    )
+    with gr.Row():
+        query_box = gr.Textbox(label="اپنا سوال لکھیں",
+                               placeholder="مثال: مصنوعی ذہانت کیا ہے؟", 
+                               scale=4, rtl=True)
+        eval_check = gr.Checkbox(label="Run Evaluation", value=True)
+        btn = gr.Button("پوچھیں", variant="primary", scale=1)
+
+    with gr.Tabs():
+        with gr.TabItem("جواب"):
+            answer_out = gr.Markdown(rtl=True)
+        with gr.TabItem("Retrieved Context"):
+            context_out = gr.Markdown(rtl=True)
+        with gr.TabItem("Evaluation Scores"):
+            scores_out = gr.Markdown()
+
+    timing_out = gr.Textbox(label="Timing", interactive=False)
+    btn.click(fn=run_query, inputs=[query_box, eval_check],
+              outputs=[answer_out, context_out, scores_out, timing_out])
+
+    gr.Markdown("---\n*Embedding: paraphrase-multilingual-MiniLM-L12-v2 | Reranker: ms-marco-MiniLM | "
+                "LLM: Mistral-7B / Llama-3-8B | Vector DB: Pinecone*")
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,6 @@
+gradio
+sentence-transformers
+rank-bm25
+pinecone
+huggingface-hub
+numpy