Add firebase logging

.gitignore

@@ -1,4 +1,4 @@
 venv/
 .venv/
 env/
-.env/
+.env

app-ngrok.py

@@ -5,15 +5,49 @@ import requests
 from time import sleep
 import re
 import platform
+# Additional Firebase imports
+import firebase_admin
+from firebase_admin import credentials, firestore
+import json
+import base64
 
 print(f"Running on {platform.system()}")
 
+if platform.system() == "Windows" or platform.system() == "Darwin":
+    from dotenv import load_dotenv
+    load_dotenv()
+
 quantized_model = "richardr1126/spider-skeleton-wizard-coder-ggml"
 merged_model = "richardr1126/spider-skeleton-wizard-coder-merged"
 initial_model = "WizardLM/WizardCoder-15B-V1.0"
 lora_model = "richardr1126/spider-skeleton-wizard-coder-qlora"
 dataset = "richardr1126/spider-skeleton-context-instruct"
 
+def log_to_firestore(input_message, db_info, temperature, response_text):
+    # For firebase
+    base64_string = os.getenv('FIREBASE')
+    base64_bytes = base64_string.encode('utf-8')
+    json_bytes = base64.b64decode(base64_bytes)
+    json_data = json_bytes.decode('utf-8')
+
+    firebase_auth = json.loads(json_data)
+
+    # Load credentials and initialize Firestore
+    cred = credentials.Certificate(firebase_auth)
+    firebase_admin.initialize_app(cred)
+    db = firestore.client()
+
+    doc_ref = db.collection('logs').document()
+    log_data = {
+        'timestamp': firestore.SERVER_TIMESTAMP,
+        'temperature': temperature,
+        'db_info': db_info,
+        'input': input_message,
+        'output': response_text
+    }
+    doc_ref.set(log_data)
+
+
 def format(text):
     # Split the text by "|", and get the last element in the list which should be the final query
     try:
@@ -33,7 +67,9 @@ def format(text):
 
     return final_query_markdown
 
-def generate(input_message: str, db_info="", temperature=0.1, top_p=0.9, top_k=0, repetition_penalty=1.08, format_sql=True, stop_sequence="###"):
+
+
+def generate(input_message: str, db_info="", temperature=0.1, top_p=0.9, top_k=0, repetition_penalty=1.08, format_sql=True, stop_sequence="###", log=False):
     # Format the user's input message
     messages = f"Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n\nConvert text to sql: {input_message} {db_info}\n\n### Response:\n\n"
 
@@ -66,10 +102,15 @@ def generate(input_message: str, db_info="", temperature=0.1, top_p=0.9, top_k=0
             if response_text and response_text[-1] == ".":
                 response_text = response_text[:-1]
 
+            # Log the request to Firestore
+            formatted_query = format(response_text)
+            log_to_firestore(input_message, db_info, temperature, formatted_query if format_sql else response_text)
+
             if format_sql:
-                return format(response_text)
+                return formatted_query
             else:
                 return response_text
+
             
         except Exception as e:
             print(f'Error occurred: {str(e)}')
@@ -102,10 +143,12 @@ with gr.Blocks(theme='gradio/soft') as demo:
     # When the button is clicked, call the generate function, inputs are taken from the UI elements, outputs are sent to outputs elements
     run_button.click(fn=generate, inputs=[input_text, db_info, temperature, top_p, top_k, repetition_penalty, format_sql, stop_sequence], outputs=output_box, api_name="txt2sql")
     
+    ## Add statement saying that inputs/outpus are sent to firebase
     info = gr.HTML(f"""
         <p>🌐 Leveraging the <a href='https://huggingface.co/{quantized_model}'><strong>4-bit GGML version</strong></a> of <a href='https://huggingface.co/{merged_model}'><strong>{merged_model}</strong></a> model.</p>
         <p>🔗 How it's made: <a href='https://huggingface.co/{initial_model}'><strong>{initial_model}</strong></a> was finetuned to create <a href='https://huggingface.co/{lora_model}'><strong>{lora_model}</strong></a>, then merged together to create <a href='https://huggingface.co/{merged_model}'><strong>{merged_model}</strong></a>.</p>
         <p>📉 Fine-tuning was performed using QLoRA techniques on the <a href='https://huggingface.co/datasets/{dataset}'><strong>{dataset}</strong></a> dataset. You can view training metrics on the <a href='https://huggingface.co/{lora_model}'><strong>QLoRa adapter HF Repo</strong></a>.</p>
+        <p>📊 All inputs/outputs are logged to Firebase, to help me see where the model still needs improvements.</a>.</p>
     """)
 
     with gr.Accordion("Examples", open=True):
@@ -135,4 +178,4 @@ with gr.Blocks(theme='gradio/soft') as demo:
             readme_content,
         )
 
-demo.queue(concurrency_count=1, max_size=10).launch(debug=True)
+demo.queue(concurrency_count=1, max_size=20).launch(debug=True)

requirements.txt

@@ -7,4 +7,5 @@ bitsandbytes
 scipy
 transformers
 accelerate
-sqlparse
+sqlparse
+firebase_admin