Add Bangla model API

app.py

@@ -1,31 +1,26 @@
 import os
 import gradio as gr
-from config import VECTOR_DIR, EMBED_MODEL, GROQ_API_KEY, GROQ_MODEL
+from config import DATA_DIR, UPLOADED_PDF_PATH, VECTOR_DIR, GROQ_API_KEY
 from ingestion import ingest_pdf
 from rag import answer, reset_memory
 
-# Prepare runtime folders
 os.makedirs(DATA_DIR, exist_ok=True)
 os.makedirs(VECTOR_DIR, exist_ok=True)
 
 def do_ingest(pdf_file):
     if GROQ_API_KEY.strip() == "":
         return "⚠️ GROQ_API_KEY is missing. Add it in your environment / HF Space secrets."
-
     if pdf_file is None:
         return "Please upload a PDF."
 
-    # Save uploaded file
     with open(UPLOADED_PDF_PATH, "wb") as f:
         f.write(pdf_file.read())
 
-    # Ingest and reset memory
     msg = ingest_pdf(UPLOADED_PDF_PATH)
     reset_memory()
     return f"✅ {msg}"
 
 def chat_fn(user_message, history):
-    # Ensure vector index exists
     if not os.path.exists(VECTOR_DIR) or not os.listdir(VECTOR_DIR):
         return "Please upload a PDF and click 'Ingest PDF' first.", history
 

config.py

@@ -1,9 +1,9 @@
 import os
 
 # ---- Paths ----
-DATA_DIR = "data"                      # Folder for uploaded PDFs
-VECTOR_DIR = "vectorstore/faiss"       # Folder for FAISS index
-UPLOADED_PDF_PATH = os.path.join(DATA_DIR, "uploaded.pdf")  # Temp path for the uploaded PDF
+DATA_DIR = "data"
+VECTOR_DIR = "vectorstore/faiss"
+UPLOADED_PDF_PATH = os.path.join(DATA_DIR, "uploaded.pdf")
 
 # ---- Models ----
 EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"

ingestion.py

@@ -6,22 +6,15 @@ from langchain_huggingface import HuggingFaceEmbeddings
 from config import VECTOR_DIR, EMBED_MODEL
 
 def ingest_pdf(pdf_path: str) -> str:
-    """
-    Load PDF, split into chunks, embed, and store in FAISS.
-    Overwrites previous FAISS index.
-    """
     if not os.path.exists(pdf_path):
         return "PDF file not found."
 
-    # Load PDF
     loader = PyPDFLoader(pdf_path)
     docs = loader.load()
 
-    # Split text
     splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
     chunks = splitter.split_documents(docs)
 
-    # Embeddings + FAISS
     embeddings = HuggingFaceEmbeddings(model_name=EMBED_MODEL)
     vectordb = FAISS.from_documents(chunks, embeddings)
     vectordb.save_local(VECTOR_DIR)

rag.py

@@ -0,0 +1,37 @@
+from langchain_community.vectorstores import FAISS
+from langchain_huggingface import HuggingFaceEmbeddings
+from langchain.memory import ConversationBufferMemory
+from langchain_groq import ChatGroq
+from langchain.chains import conversational_retrieval
+from langchain.prompts import ChatPromptTemplate, SystemMessagePromptTemplate, HumanMessagePromptTemplate
+from config import VECTOR_DIR, EMBED_MODEL, GROQ_API_KEY, GROQ_MODEL
+
+_memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
+
+def reset_memory():
+    _memory.clear()
+
+def build_chain():
+    embeddings = HuggingFaceEmbeddings(model_name=EMBED_MODEL)
+    vectordb = FAISS.load_local(VECTOR_DIR, embeddings, allow_dangerous_deserialization=True)
+    retriever = vectordb.as_retriever(search_kwargs={"k": 3})
+
+    llm = ChatGroq(model=GROQ_MODEL, api_key=GROQ_API_KEY, temperature=0.1)
+
+    chat_prompt = ChatPromptTemplate.from_messages([
+        SystemMessagePromptTemplate.from_template("You are a helpful assistant. Answer using the context."),
+        HumanMessagePromptTemplate.from_template("Context:\n{context}\n\nQuestion:\n{question}\nAnswer clearly and concisely:")
+    ])
+
+    chain = conversational_retrieval.from_llm(
+        llm=llm,
+        retriever=retriever,
+        memory=_memory,
+        combine_docs_chain_kwargs={"prompt": chat_prompt}
+    )
+    return chain
+
+def answer(question: str) -> str:
+    chain = build_chain()
+    result = chain.invoke({"question": question})
+    return result.get("answer") or result.get("result") or "No answer."