Load and process the knowledge.txt file

app.py

@@ -7,6 +7,42 @@ import numpy as np
 # this client will handle making requests to the model to generate responses
 client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
 
+# Load and process the knowledge base text file
+with open("knowledge.txt", "r", encoding="utf-8") as f:
+    knowledge_text = f.read()
+
+# Split the text into chunks (for example, by paragraphs)
+chunks = [chunk.strip() for chunk in knowledge_text.split("\n\n") if chunk.strip()]
+
+# Load an embedding model (this one is light and fast)
+embedder = SentenceTransformer('all-MiniLM-L6-v2')
+
+# Precompute embeddings for all chunks (as a tensor for fast similarity search)
+chunk_embeddings = embedder.encode(chunks, convert_to_tensor=True)
+
+def get_relevant_context(query, top_k=3):
+    """
+     Compute the embedding for the query, compare it against all chunk embeddings,
+     and return the top_k most similar chunks concatenated into a context string.
+    """
+    
+    # Compute and normalize the query embedding
+    query_embedding = embedder.encode(query, convert_to_tensor=True)
+    query_embedding = query_embedding / query_embedding.norm()
+    
+    # Normalize chunk embeddings along the embedding dimension
+    norm_chunk_embeddings = chunk_embeddings / chunk_embeddings.norm(dim=1, keepdim=True)
+    
+    # Compute cosine similarity between the query and each chunk
+    similarities = torch.matmul(norm_chunk_embeddings, query_embedding)
+    
+    # Get the indices of the top_k most similar chunks
+    top_k_indices = torch.topk(similarities, k=top_k).indices.cpu().numpy()
+    
+    # Concatenate the top chunks into a single context string
+    context = "\n\n".join([chunks[i] for i in top_k_indices])
+    return context
+
 def respond(message, history):
 
     system_message = "You are a kitchen sous chef. You always respond with a knowledgeable and upbeat attitude!"
@@ -66,7 +102,7 @@ with gr.Blocks() as chatbot:
         with gr.Column(scale=2):
             gr.ChatInterface(respond,
                              type="messages",
-                             examples = ["How do I know if my chicken is fully cooked?", "What toppings should I add to my ramen?", "What's the best recipe for homemade pizza dough?"],
+                             examples = ["What are some cooking tips for beginners?", "What toppings should I add to my ramen?", "What's the best recipe for homemade pizza dough?"],
                              theme='d8ahazard/material_design_rd'
                             )
     with gr.Row():