Session 15: filled in blanks

app.py

@@ -16,7 +16,7 @@ import asyncio
 from tqdm.asyncio import tqdm
 
 # GLOBAL SCOPE - ENTIRE APPLICATION HAS ACCESS TO VALUES SET IN THIS SCOPE #
-# ---- ENV VARIABLES ---- # 
+# ---- ENV VARIABLES ---- #
 """
 This function will load our environment file (.env) if it is present.
 
@@ -41,20 +41,26 @@ HF_TOKEN = os.environ["HF_TOKEN"]
 4. Index Files if they do not exist, otherwise load the vectorstore
 """
 ### 1. CREATE TEXT LOADER AND LOAD DOCUMENTS
-### NOTE: PAY ATTENTION TO THE PATH THEY ARE IN. 
-text_loader = 
-documents = 
+### NOTE: PAY ATTENTION TO THE PATH THEY ARE IN.
+text_loader = TextLoader("./data/paul_graham_essays.txt")
+documents = text_loader.load()
 
 ### 2. CREATE TEXT SPLITTER AND SPLIT DOCUMENTS
-text_splitter = 
-split_documents = 
+text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=30)
+split_documents = text_splitter.split_documents(documents)
 
 ### 3. LOAD HUGGINGFACE EMBEDDINGS
-hf_embeddings = 
+hf_embeddings = HuggingFaceEndpointEmbeddings(
+    model=HF_EMBED_ENDPOINT,
+    task="feature-extraction",
+    huggingfacehub_api_token=HF_TOKEN,
+)
+
 
 async def add_documents_async(vectorstore, documents):
     await vectorstore.aadd_documents(documents)
 
+
 async def process_batch(vectorstore, batch, is_first_batch, pbar):
     if is_first_batch:
         result = await FAISS.afrom_documents(batch, hf_embeddings)
@@ -64,44 +70,52 @@ async def process_batch(vectorstore, batch, is_first_batch, pbar):
     pbar.update(len(batch))
     return result
 
+
 async def main():
     print("Indexing Files")
-    
+
     vectorstore = None
     batch_size = 32
-    
-    batches = [split_documents[i:i+batch_size] for i in range(0, len(split_documents), batch_size)]
-    
+
+    batches = [
+        split_documents[i : i + batch_size]
+        for i in range(0, len(split_documents), batch_size)
+    ]
+
     async def process_all_batches():
         nonlocal vectorstore
         tasks = []
         pbars = []
-        
+
         for i, batch in enumerate(batches):
-            pbar = tqdm(total=len(batch), desc=f"Batch {i+1}/{len(batches)}", position=i)
+            pbar = tqdm(
+                total=len(batch), desc=f"Batch {i+1}/{len(batches)}", position=i
+            )
             pbars.append(pbar)
-            
-            if i == 0:
+
+            if i == 0:  # first batch is processed directly to initialize vectorstore
                 vectorstore = await process_batch(None, batch, True, pbar)
-            else:
+            else:  # the remaining batches are processed in parallel
                 tasks.append(process_batch(vectorstore, batch, False, pbar))
-        
+
         if tasks:
             await asyncio.gather(*tasks)
-        
+
         for pbar in pbars:
             pbar.close()
-    
+
     await process_all_batches()
-    
+
     hf_retriever = vectorstore.as_retriever()
     print("\nIndexing complete. Vectorstore is ready for use.")
     return hf_retriever
 
+
 async def run():
     retriever = await main()
     return retriever
 
+
 hf_retriever = asyncio.run(run())
 
 # -- AUGMENTED -- #
@@ -110,46 +124,72 @@ hf_retriever = asyncio.run(run())
 2. Create a Prompt Template from the String Template
 """
 ### 1. DEFINE STRING TEMPLATE
-RAG_PROMPT_TEMPLATE = 
+RAG_PROMPT_TEMPLATE = """\
+<|start_header_id|>system<|end_header_id|>
+You are a helpful assistant. You answer user questions based on provided context. If you can't answer the question with the provided context, say you don't know.<|eot_id|>
+
+<|start_header_id|>user<|end_header_id|>
+User Query:
+{query}
+
+Context:
+{context}<|eot_id|>
+
+<|start_header_id|>assistant<|end_header_id|>
+"""
 
 ### 2. CREATE PROMPT TEMPLATE
-rag_prompt =
+rag_prompt = PromptTemplate.from_template(RAG_PROMPT_TEMPLATE)
 
 # -- GENERATION -- #
 """
 1. Create a HuggingFaceEndpoint for the LLM
 """
 ### 1. CREATE HUGGINGFACE ENDPOINT FOR LLM
-hf_llm = 
+hf_llm = HuggingFaceEndpoint(
+    endpoint_url=f"{HF_LLM_ENDPOINT}",
+    task="text-generation",
+    max_new_tokens=512,
+    top_k=10,
+    top_p=0.95,
+    typical_p=0.95,
+    temperature=0.01,
+    repetition_penalty=1.03,
+)
+
 
 @cl.author_rename
 def rename(original_author: str):
     """
-    This function can be used to rename the 'author' of a message. 
+    This function can be used to rename the 'author' of a message.
 
     In this case, we're overriding the 'Assistant' author to be 'Paul Graham Essay Bot'.
     """
-    rename_dict = {
-        "Assistant" : "Paul Graham Essay Bot"
-    }
+    rename_dict = {"Assistant": "Paul Graham Essay Bot"}
     return rename_dict.get(original_author, original_author)
 
+
 @cl.on_chat_start
 async def start_chat():
     """
-    This function will be called at the start of every user session. 
+    This function will be called at the start of every user session.
 
-    We will build our LCEL RAG chain here, and store it in the user session. 
+    We will build our LCEL RAG chain here, and store it in the user session.
 
     The user session is a dictionary that is unique to each user session, and is stored in the memory of the server.
     """
 
     ### BUILD LCEL RAG CHAIN THAT ONLY RETURNS TEXT
-    lcel_rag_chain = 
+    lcel_rag_chain = (
+        {"context": itemgetter("query") | hf_retriever, "query": itemgetter("query")}
+        | rag_prompt
+        | hf_llm
+    )
 
     cl.user_session.set("lcel_rag_chain", lcel_rag_chain)
 
-@cl.on_message  
+
+@cl.on_message
 async def main(message: cl.Message):
     """
     This function will be called every time a message is recieved from a session.
@@ -168,4 +208,4 @@ async def main(message: cl.Message):
     ):
         await msg.stream_token(chunk)
 
-    await msg.send()
+    await msg.send()