Update app.py

✔ No more crashes from 429
The retry logic absorbs rate limits gracefully.
✔ Gradio won’t spawn multiple parallel LLM calls
concurrency_limit=1 prevents overload.

app.py

@@ -1,10 +1,12 @@
 import os
+import time
 import gradio as gr
 import numpy as np
 import faiss
 from PyPDF2 import PdfReader
 from sentence_transformers import SentenceTransformer
 from huggingface_hub import InferenceClient
+from huggingface_hub.errors import HfHubHTTPError
 
 # ============================================================
 # CONFIG
@@ -12,27 +14,36 @@ from huggingface_hub import InferenceClient
 
 HF_TOKEN = os.getenv("HF_TOKEN", None)
 
-# Free model compatible with new HF API
-#LLM_MODEL = "meta-llama/Llama-3.2-1B-Instruct"
-#LLM_MODEL = "meta-llama/Llama-3.1-8B-Instruct"
-# LLM_MODEL = "mistralai/Mistral-7B-Instruct-v0.3" Does not work
-# LLM_MODEL ="HuggingFaceH4/zephyr-7b-beta"
-# LLM_MODEL = "NousResearch/Hermes-2-Pro-Mistral-7B" Does not work
-
-# Great LLM
-#LLM_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
+# Recommended: 7B is more stable on free HF router
+# LLM_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
 LLM_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
 
-# Local embedding model
 EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
 
-# Load models
 embedding_model = SentenceTransformer(EMBEDDING_MODEL)
 
 # IMPORTANT: do NOT pass model=... here
 llm = InferenceClient(token=HF_TOKEN)
 
 
+# ============================================================
+# SAFE LLM CALL (prevents 429 errors)
+# ============================================================
+
+def safe_chat_completion(llm, **kwargs):
+    for attempt in range(5):
+        try:
+            return llm.chat.completions.create(**kwargs)
+        except HfHubHTTPError as e:
+            if "429" in str(e):
+                sleep_time = 2 ** attempt
+                print(f"[Rate limit] Retrying in {sleep_time}s...")
+                time.sleep(sleep_time)
+            else:
+                raise
+    raise RuntimeError("Too many retries after rate limits.")
+
+
 # ============================================================
 # PDF PROCESSING
 # ============================================================
@@ -89,7 +100,7 @@ def retrieve_context(question, state, k=4):
 
 
 # ============================================================
-# LLM ANSWERING (NEW HF API)
+# LLM ANSWERING
 # ============================================================
 
 def generate_answer(question, context):
@@ -106,14 +117,15 @@ def generate_answer(question, context):
             f"Question: {question}\nAnswer:"
         )
 
-    response = llm.chat.completions.create(
+    response = safe_chat_completion(
+        llm,
         model=LLM_MODEL,
         messages=[{"role": "user", "content": prompt}],
-        max_tokens=2048,
+        max_tokens=1024,
         temperature=0.25,
         top_p=0.9,
-        frequency_penalty=0.1,  # reduces repetition
-        presence_penalty=0.0    # keeps answers grounded
+        frequency_penalty=0.1,
+        presence_penalty=0.0
     )
 
     return response.choices[0].message["content"].strip()
@@ -164,7 +176,7 @@ def answer_question(question, state):
 # GRADIO UI
 # ============================================================
 
-with gr.Blocks(title="PDF Q&A Chatbot") as demo:
+with gr.Blocks(title="PDF Q&A Chatbot", concurrency_limit=1) as demo:
     gr.Markdown("## 📄 PDF Q&A Chatbot\nUpload a PDF and ask questions about its content.")
 
     state = gr.State({"chunks": None, "index": None})