ADDED: Advance RAG System

.gitignore

@@ -3,4 +3,5 @@ venv/
 .env
 .pytest_cache/
 .vscode
-hit.py
+hit.py
+try.ipynb

README.md

@@ -151,7 +151,7 @@ percentage, missing keywords, and profile summary.
 ### 19. Text to Image using Diffusion Models
 
 - **Route:** `/text2image`
-description: This route allows you to generate images using various diffusion models available on Hugging Face.
+- **Description:** This route allows you to generate images using various diffusion models available on Hugging Face.
 
     You can choose from the following models -
 
@@ -167,6 +167,15 @@ description: This route allows you to generate images using various diffusion mo
 
     **Request Body:** `{ "model": "DreamShaper_v7", "prompt": "An astronaut riding a horse on the moon" }` The response will be the generated image in PNG format.
 
+### 20. Advanced RAG with LLaMA Index
+
+- **Route:** `/advance_rag_llama_index`
+- **Description:** This API provides an advanced retrieval-augmented generation (RAG) functionality using the LLaMA Index. It allows users to upload a document and ask questions, enabling the API to search for specific facts or summarize the document based on the query.
+- **Feature:**
+  - Upload PDF documents and ask natural language questions.
+  - Choose between vector search or summarization for responses.
+  - Support for multiple Open Source LLMs models.
+
 ## Usage
 
 Each endpoint accepts specific parameters as described in the respective endpoint documentation. Users can make POST requests to these endpoints with the required parameters to perform the desired tasks.

api.py

@@ -15,7 +15,7 @@
 from mongo import MongoDB
 from helper_functions import get_qa_chain,get_gemini_response,get_url_doc_qa,extract_transcript_details,\
     get_gemini_response_health,get_gemini_pdf,read_sql_query,remove_substrings,questions_generator,groq_pdf,\
-    summarize_audio,chatbot_send_message,extraxt_pdf_text
+    summarize_audio,chatbot_send_message,extraxt_pdf_text,advance_rag_llama_index
 from langchain_groq import ChatGroq
 from langchain.chains import ConversationChain
 from langchain.chains.conversation.memory import ConversationBufferWindowMemory
@@ -529,7 +529,27 @@ async def ats(resume_pdf: UploadFile = File(...), job_description: str = Form(..
         return ResponseText(response=response.text)
     except Exception as e:
         return ResponseText(response=f"Error: {str(e)}")
-
+
+@app.post("/advance_rag_llama_index",description="The endpoint build a Router that can choose whether to do vector search or summarization\
+          In model input default is llama3-70b-8192 but you can choose mixtral-8x7b-32768, gemma-7b-it and llama3-8b-8192.")
+async def llama_index_rag(pdf: UploadFile = File(...),question: str = Form(...),
+                       model: Optional[str] = Form('llama3-70b-8192')):
+    try:
+        rag_output = advance_rag_llama_index(pdf,model,question)
+        db = MongoDB()
+        payload = {
+            "endpoint" : "/advance_rag_llama_index",
+            "model" : model,
+            "prompt" : question,
+            "output" : rag_output
+        }
+        mongo_data = {"Document": payload}
+        result = db.insert_data(mongo_data)
+        print(result)
+        return ResponseText(response=rag_output)
+    except Exception as e:
+        return ResponseText(response=f"Error: {str(e)}")
+
 @app.post("/text2image",description=
         """
             This API provides access to the following diffusion models for generating images from text prompts.

helper_functions.py

@@ -24,6 +24,13 @@
 from typing import AsyncIterable
 import asyncio
 from langchain.schema import HumanMessage
+from llama_index.llms.groq import Groq
+from langchain_community.embeddings import HuggingFaceInferenceAPIEmbeddings
+from llama_index.core import SimpleDirectoryReader,VectorStoreIndex
+from llama_index import core
+from llama_index.core.tools import QueryEngineTool, ToolMetadata
+from llama_index.core.query_engine import RouterQueryEngine
+import shutil
 
 load_dotenv()
 genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))
@@ -272,6 +279,46 @@ def extraxt_pdf_text(uploaded_file):
         text+=str(page.extract_text())
     return text
 
+def advance_rag_llama_index(pdf,model,question):
+    try:
+        with tempfile.NamedTemporaryFile(delete=False, suffix='.' + pdf.filename.split('.')[-1]) as tmp:
+            shutil.copyfileobj(pdf.file, tmp)
+            tmp_path = tmp.name
+    except Exception as e:
+        raise Exception(f"Error handling uploaded file: {e}")
+
+    finally:
+        pdf.file.close()
+    llm = Groq(model=model,api_key=os.getenv("GROQ_API_KEY"))
+    embed_model = HuggingFaceInferenceAPIEmbeddings(
+    api_key=os.getenv("HUGGINGFACE_API_KEY"), model_name=settings.INSTRUCTOR_EMBEDDING,query_instruction="Represent the query for retrieval: ")
+
+    core.Settings.llm = llm
+    core.Settings.embed_model = embed_model
+
+    docs = SimpleDirectoryReader(input_files=[tmp_path]).load_data()
+    index = VectorStoreIndex.from_documents(docs)
+    vector_tool = QueryEngineTool(
+    index.as_query_engine(),
+    metadata=ToolMetadata(
+        name="vector_search",
+        description="Useful for searching for specific facts."))
+
+    summary_tool = QueryEngineTool(
+    index.as_query_engine(response_mode="tree_summarize"),
+    metadata=ToolMetadata(
+        name="summary",
+        description="Useful for summarizing an entire document."))
+
+    query_engine = RouterQueryEngine.from_defaults(
+        [vector_tool, summary_tool], select_multi=False, verbose=True, llm=llm)
+
+    response = query_engine.query(question)
+    os.remove(tmp_path)
+
+    return str(response)
+
+
 if __name__ == "__main__":
     create_vector_db()
     chain = get_qa_chain()

requirements.txt

@@ -20,4 +20,6 @@ jinja2
 tiktoken
 redis
 python-jose
-sendgrid
+sendgrid
+llama-index-llms-groq
+llama-index-embeddings-langchain