Update app.py

app.py

@@ -1,64 +1,174 @@
+# app.py
+import os, io, re, json
 import gradio as gr
-from huggingface_hub import InferenceClient
-
-"""
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
-
-
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
-
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
-
-    messages.append({"role": "user", "content": message})
-
-    response = ""
-
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
-
-        response += token
-        yield response
-
-
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
+from pdfminer.high_level import extract_text
+from sentence_transformers import SentenceTransformer
+from qdrant_client import QdrantClient
+from qdrant_client.http import models as qmodels
+import requests
 
+# ------------ Global (lazy) ------------
+EMBED_MODEL_ID = "intfloat/multilingual-e5-small"  # 384-dim, multilingual
+_embedder = None
+_client = None
+_collection = None
+
+# ------------ Helpers ------------
+def get_embedder():
+    global _embedder
+    if _embedder is None:
+        _embedder = SentenceTransformer(EMBED_MODEL_ID)
+    return _embedder
+
+def connect_qdrant(url: str, api_key: str, collection: str):
+    global _client, _collection
+    _client = QdrantClient(url=url, api_key=api_key) if url else None
+    _collection = collection
+    if _client:
+        try:
+            _client.get_collection(collection)
+        except Exception:
+            _client.recreate_collection(
+                collection_name=collection,
+                vectors_config=qmodels.VectorParams(size=384, distance=qmodels.Distance.COSINE)
+            )
+    return True
+
+def chunk_text(text: str, size: int = 1500, overlap: int = 200):
+    text = re.sub(r"\s+", " ", text)
+    chunks, i = [], 0
+    step = max(1, size - overlap)
+    while i < len(text):
+        chunk = text[i:i+size]
+        if chunk.strip():
+            chunks.append(chunk)
+        i += step
+    return chunks
+
+def embed_passages(texts):
+    emb = get_embedder()
+    return emb.encode([f"passage: {t}" for t in texts], normalize_embeddings=True).tolist()
+
+def embed_query(text: str):
+    emb = get_embedder()
+    return emb.encode([f"query: {text}"], normalize_embeddings=True)[0].tolist()
+
+# ------------ Ingest ------------
+def ingest_pdfs(files, topic: str, chunk_chars: int, overlap: int):
+    if not _client:
+        return "❌ Najprv nastav QDRANT_URL a QDRANT_API_KEY vľavo.", None
+    total_chunks = 0
+    for f in files or []:
+        data = f.read()
+        text = extract_text(io.BytesIO(data))
+        chunks = chunk_text(text, chunk_chars, overlap)
+        vecs = embed_passages(chunks)
+        points = []
+        for c, v in zip(chunks, vecs):
+            points.append(qmodels.PointStruct(vector=v, payload={"text": c, "topic": topic}))
+        _client.upsert(collection_name=_collection, points=points)
+        total_chunks += len(chunks)
+    return f"✅ Uložené chunky: {total_chunks}", None
+
+# ------------ Search & (optional) web ------------
+def search_local(query: str, topk: int, topic: str):
+    if not _client:
+        return []
+    qvec = embed_query(query)
+    qfilter = qmodels.Filter(must=[qmodels.FieldCondition(key="topic", match=qmodels.MatchValue(value=topic))]) if topic else None
+    res = _client.search(collection_name=_collection, query_vector=qvec, limit=topk, query_filter=qfilter)
+    return [{"score": r.score, "text": r.payload.get("text",""), "payload": r.payload} for r in res]
+
+def web_search(query: str, n: int = 3, api_key: str | None = None):
+    key = api_key or os.getenv("TAVILY_API_KEY", "")
+    if not key:
+        return []
+    r = requests.post("https://api.tavily.com/search", json={
+        "api_key": key,
+        "query": query,
+        "search_depth": "basic",
+        "max_results": n
+    }, timeout=30)
+    data = r.json()
+    out = []
+    for item in data.get("results", [])[:n]:
+        text = (item.get("title","") + "\n" + item.get("content",""))[:1500]
+        out.append({"url": item.get("url"), "text": text})
+    return out
+
+# ------------ Q&A without LLM (template-based) ------------
+def answer_from_context(question: str, local_hits, web_snips):
+    parts = ["**Otázka:** " + question, "\n**Odpoveď (kompilované zo zdrojov):**"]
+    if local_hits:
+        parts.append("- Z lokálnych PDF vyplýva:")
+        for i,h in enumerate(local_hits):
+            snippet = h['text'][:400].strip().replace('\n',' ')
+            parts.append(f"  [{i+1}] {snippet} … (score {h['score']:.2f})")
+    if web_snips:
+        parts.append("- Z webu:")
+        for j,w in enumerate(web_snips):
+            snippet = w['text'][:300].strip().replace('\n',' ')
+            parts.append(f"  [W{j+1}] {snippet} … (zdroj: {w['url']})")
+    parts.append("\n**Poznámka:** Toto je súhrn citovaných pasáží. Pre presnejšiu odpoveď pridaj viac PDF alebo zapni lepší model neskôr.")
+    return "\n".join(parts)
+
+# ------------ Gradio UI ------------
+def on_connect(url, key, collection):
+    try:
+        connect_qdrant(url, key, collection)
+        return gr.update(value=f"✅ Pripojené k Qdrant: {collection}"), ""
+    except Exception as e:
+        return gr.update(value=f"❌ Chyba pripojenia"), str(e)
+
+with gr.Blocks(title="EDM & Fanuc – Hybrid RAG (PDF + Web)") as demo:
+    gr.Markdown("# 🔧 EDM & Fanuc – Hybrid RAG (PDF + Web)")
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            gr.Markdown("### Qdrant nastavenie")
+            q_url = gr.Textbox(label="QDRANT_URL", value=os.getenv("QDRANT_URL", ""))
+            q_key = gr.Textbox(label="QDRANT_API_KEY", value=os.getenv("QDRANT_API_KEY", ""), type="password")
+            collection_tb = gr.Textbox(label="Collection", value="docs")
+            connect_btn = gr.Button("🔌 Connect")
+            status = gr.Textbox(label="Status", interactive=False)
+            err = gr.Textbox(label="Error", interactive=False)
+
+            def _connect(u,k,c):
+                s,e = on_connect(u,k,c)
+                return s,e
+            connect_btn.click(_connect, inputs=[q_url,q_key,collection_tb], outputs=[status, err])
+
+            gr.Markdown("### Ingest PDF")
+            topic_dd = gr.Dropdown(["EDM","FANUC","general"], value="EDM", label="Topic")
+            chunk_chars = gr.Slider(800, 2500, value=1500, step=100, label="Chunk size")
+            overlap = gr.Slider(0, 400, value=200, step=20, label="Overlap")
+            upload = gr.File(file_count="multiple", file_types=[".pdf"], label="PDF súbory")
+            ingest_btn = gr.Button("📥 Indexovať do Qdrant")
+            ingest_out = gr.Textbox(label="Výstup")
+
+            def _ingest(files, topic, size, ov):
+                msg,_ = ingest_pdfs(files, topic, size, ov)
+                return msg
+            ingest_btn.click(_ingest, inputs=[upload, topic_dd, chunk_chars, overlap], outputs=ingest_out)
+
+        with gr.Column(scale=2):
+            gr.Markdown("### Chat (PDF + voliteľne web)")
+            question = gr.Textbox(label="Otázka", placeholder="Napr. G41 vs G42 na Fanuc…", lines=3)
+            topic_chat = gr.Dropdown(["EDM","FANUC","general"], value="EDM", label="Topic filter")
+            topk = gr.Slider(1, 8, value=5, step=1, label="Top‑K")
+            minscore = gr.Slider(0, 1, value=0.35, step=0.01, label="Min lokálne score pre vypnutie webu")
+            use_web = gr.Checkbox(value=True, label="Povoliť web doplnenie (Tavily)")
+            tavily_key = gr.Textbox(label="TAVILY_API_KEY (nepovinné)", value=os.getenv("TAVILY_API_KEY",""), type="password")
+            ask_btn = gr.Button("💬 Spýtať sa")
+            answer_md = gr.Markdown()
+
+            def _ask(q, topic, k, thr, web, tkey):
+                local_hits = search_local(q, int(k), topic)
+                good = any(h.get("score",0) >= float(thr) for h in local_hits)
+                web_snips = web_search(q, n=3, api_key=tkey) if (web and not good) else []
+                return answer_from_context(q, local_hits, web_snips)
+
+            ask_btn.click(_ask, inputs=[question, topic_chat, topk, minscore, use_web, tavily_key], outputs=answer_md)
 
 if __name__ == "__main__":
     demo.launch()