Update app.py

app.py

@@ -2,10 +2,9 @@ import os
 import time
 import gradio as gr
 import numpy as np
+import requests
 from PyPDF2 import PdfReader
 from sentence_transformers import SentenceTransformer
-from huggingface_hub import InferenceClient
-from huggingface_hub.errors import HfHubHTTPError
 
 # ============================================================
 # CONFIG
@@ -13,44 +12,32 @@ from huggingface_hub.errors import HfHubHTTPError
 
 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
-
-# Were Great LLM
-# LLM_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
-# LLM_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
-
-# LLM model
-# LLM_MODEL = "deepseek-ai/DeepSeek-V3"
-LLM_MODEL = 'Qwen/Qwen2.5-14B-Instruct'
+# FREE MODEL (no credits required)
+FREE_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
 
 EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
 embedding_model = SentenceTransformer(EMBEDDING_MODEL)
 
-# IMPORTANT: do NOT pass model=... here
-llm = InferenceClient(token=HF_TOKEN)
-
 
 # ============================================================
-# SAFE LLM CALL (prevents 429 errors)
+# FREE HF INFERENCE API (NO CREDITS REQUIRED)
 # ============================================================
 
-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.")
+def hf_free_inference(model, prompt):
+    response = requests.post(
+        f"https://api-inference.huggingface.co/models/{model}",
+        headers={"Authorization": f"Bearer {HF_TOKEN}"},
+        json={"inputs": prompt}
+    )
+    data = response.json()
+
+    # Some models return a list, others return a dict
+    if isinstance(data, list) and "generated_text" in data[0]:
+        return data[0]["generated_text"]
+    if isinstance(data, dict) and "generated_text" in data:
+        return data["generated_text"]
+
+    return str(data)
 
 
 # ============================================================
@@ -86,10 +73,6 @@ def chunk_text(text, chunk_size=800, overlap=200):
 # ============================================================
 
 def retrieve_context_per_file(question, state, k=10):
-    """
-    For each file, retrieve top-k chunks separately,
-    then merge all selected chunks into a single context.
-    """
     if not state["files"]:
         return ""
 
@@ -103,13 +86,9 @@ def retrieve_context_per_file(question, state, k=10):
         if len(chunks) == 0:
             continue
 
-        # Compute L2 distances to the query embedding
         distances = np.linalg.norm(embeddings - q_emb, axis=1)
-
-        # Get top-k indices for this file
         top_k_idx = distances.argsort()[: min(k, len(chunks))]
 
-        # Tag chunks with filename for transparency
         selected = [f"[{filename}]\n{chunks[i]}" for i in top_k_idx]
         all_selected.extend(selected)
 
@@ -137,18 +116,8 @@ def generate_answer(question, context):
             f"Question: {question}\nAnswer:"
         )
 
-    response = safe_chat_completion(
-        llm,
-        model=LLM_MODEL,
-        messages=[{"role": "user", "content": prompt}],
-        max_tokens=1024,
-        temperature=0.25,
-        top_p=0.9,
-        frequency_penalty=0.1,
-        presence_penalty=0.0,
-    )
-
-    return response.choices[0].message["content"].strip()
+    answer = hf_free_inference(FREE_MODEL, prompt)
+    return answer.strip()
 
 
 # ============================================================
@@ -156,25 +125,6 @@ def generate_answer(question, context):
 # ============================================================
 
 def load_pdf(files, state):
-    """
-    files: current list from the UI (after add/remove).
-    We rebuild state from scratch to stay in sync.
-
-    state = {
-        "files": {
-            "filename.pdf": {
-                "chunks": [...],
-                "embeddings": np.array([...])
-            },
-            ...
-        },
-        "pdf_count": int,
-        "total_chunks": int,
-        "history": [...],
-        "facts": [...]
-    }
-    """
-    # If user cleared all files
     if not files:
         state["files"] = {}
         state["pdf_count"] = 0
@@ -187,12 +137,10 @@ def load_pdf(files, state):
         )
 
     try:
-        # Reset and rebuild from current files only
         state["files"] = {}
         state["pdf_count"] = 0
         state["total_chunks"] = 0
 
-        # Deduplicate while preserving order
         unique_files = list(dict.fromkeys(files))
 
         for file in unique_files:
@@ -241,10 +189,8 @@ def remove_pdf(filename, state):
             gr.update(choices=list(state["files"].keys()), value=None),
         )
 
-    # Remove the file entry
     state["files"].pop(filename, None)
 
-    # Recompute counts
     pdf_count = len(state["files"])
     total_chunks = sum(len(data["chunks"]) for data in state["files"].values())
 
@@ -264,14 +210,6 @@ def remove_pdf(filename, state):
 
 
 def answer_question(question, state):
-    print("HF_TOKEN loaded:", HF_TOKEN is not None)
-
-    if HF_TOKEN:
-        print("HF_TOKEN length:", len(HF_TOKEN))
-        print("HF_TOKEN preview:", HF_TOKEN[:6] + "..." + HF_TOKEN[-4:])
-    else:
-        print("HF_TOKEN is None or empty")
-
     question = question.strip()
     if not question:
         return "Please enter a question."
@@ -279,7 +217,7 @@ def answer_question(question, state):
     if not state["files"]:
         return "Please upload at least one PDF first."
 
-    # Simple persistent memory hook: "Remember that ..."
+    # Persistent memory: "Remember that ..."
     lower_q = question.lower()
     if lower_q.startswith("remember that"):
         fact = question[len("remember that"):].strip(" .")
@@ -289,10 +227,8 @@ def answer_question(question, state):
         else:
             return "What should I remember?"
 
-    # RAG context (per-file balanced retrieval)
     rag_context = retrieve_context_per_file(question, state, k=10)
 
-    # Conversation memory (last 5 turns)
     memory_context = ""
     if state["history"]:
         memory_context = "\n\n".join(
@@ -302,12 +238,10 @@ def answer_question(question, state):
             ]
         )
 
-    # Persistent facts
     facts_context = ""
     if state["facts"]:
         facts_context = "User Facts:\n" + "\n".join(state["facts"]) + "\n\n"
 
-    # Merge all contexts
     full_context_parts = []
     if facts_context:
         full_context_parts.append(facts_context)
@@ -318,10 +252,8 @@ def answer_question(question, state):
 
     full_context = "\n\n".join(full_context_parts)
 
-    # Generate answer
     answer = generate_answer(question, full_context)
 
-    # Save conversational memory
     state["history"].append({"user": question, "assistant": answer})
 
     return answer
@@ -336,21 +268,20 @@ with gr.Blocks(title="PDF Q&A Chatbot") as demo:
 
     state = gr.State(
         {
-            "files": {},          # filename -> {chunks, embeddings}
+            "files": {},
             "pdf_count": 0,
             "total_chunks": 0,
-            "history": [],        # conversational memory
-            "facts": [],          # persistent user facts
+            "history": [],
+            "facts": [],
         }
     )
 
     with gr.Row():
-        # LEFT COLUMN — PDF upload + removal
         with gr.Column(scale=1):
             pdf_input = gr.File(
                 label="Upload PDF(s)",
                 file_types=[".pdf"],
-                file_count="multiple",  # proper multi-upload in Gradio 6
+                file_count="multiple",
             )
             pdf_status = gr.Textbox(label="Status", interactive=False)
 
@@ -361,7 +292,6 @@ with gr.Blocks(title="PDF Q&A Chatbot") as demo:
             )
             remove_button = gr.Button("Remove Selected PDF")
 
-        # RIGHT COLUMN — question + clear + answer
         with gr.Column(scale=2):
             with gr.Row():
                 question = gr.Textbox(
@@ -379,7 +309,6 @@ with gr.Blocks(title="PDF Q&A Chatbot") as demo:
             answer = gr.Textbox(label="Answer", lines=8)
             ask_button = gr.Button("Ask")
 
-    # Bind events AFTER components exist
     pdf_input.change(
         load_pdf,
         inputs=[pdf_input, state],