PhiRAG / psyllm.py
alx-d's picture
Upload folder using huggingface_hub
f4368a7 verified
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
import datetime
import functools
import traceback
from typing import List, Optional, Any, Dict
import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from langchain_community.llms import HuggingFacePipeline
# Other LangChain and community imports
from langchain_community.document_loaders import OnlinePDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain.embeddings import HuggingFaceEmbeddings
from langchain_community.retrievers import BM25Retriever
from langchain.retrievers import EnsembleRetriever
from langchain.prompts import ChatPromptTemplate
from langchain.schema import StrOutputParser, Document
from langchain_core.runnables import RunnableParallel, RunnableLambda
from transformers.quantizers.auto import AutoQuantizationConfig
import gradio as gr
import requests
from pydantic import PrivateAttr
import pydantic
from langchain.llms.base import LLM
from typing import Any, Optional, List
import typing
import time
print("Pydantic Version: ")
print(pydantic.__version__)
# Add Mistral imports with fallback handling
try:
from mistralai import Mistral
MISTRAL_AVAILABLE = True
debug_print = lambda msg: print(f"[{datetime.datetime.now().isoformat()}] {msg}")
debug_print("Loaded latest Mistral client library")
except ImportError:
MISTRAL_AVAILABLE = False
debug_print = lambda msg: print(f"[{datetime.datetime.now().isoformat()}] {msg}")
debug_print("Mistral client library not found. Install with: pip install mistralai")
def debug_print(message: str):
print(f"[{datetime.datetime.now().isoformat()}] {message}", flush=True)
def word_count(text: str) -> int:
return len(text.split())
# Initialize a tokenizer for token counting (using gpt2 as a generic fallback)
def initialize_tokenizer():
try:
return AutoTokenizer.from_pretrained("gpt2")
except Exception as e:
debug_print("Failed to initialize tokenizer: " + str(e))
return None
global_tokenizer = initialize_tokenizer()
def count_tokens(text: str) -> int:
if global_tokenizer:
try:
return len(global_tokenizer.encode(text))
except Exception as e:
return len(text.split())
return len(text.split())
# Add these imports at the top of your file
import uuid
import threading
import queue
from typing import Dict, Any, Tuple, Optional
import time
# Global storage for jobs and results
jobs = {} # Stores job status and results
results_queue = queue.Queue() # Thread-safe queue for completed jobs
processing_lock = threading.Lock() # Prevent simultaneous processing of the same job
# Add a global variable to store the last job ID
last_job_id = None
# Add these missing async processing functions
def process_in_background(job_id, function, args):
"""Process a function in the background and store results"""
try:
debug_print(f"Processing job {job_id} in background")
result = function(*args)
results_queue.put((job_id, result))
debug_print(f"Job {job_id} completed and added to results queue")
except Exception as e:
debug_print(f"Error in background job {job_id}: {str(e)}")
error_result = (f"Error processing job: {str(e)}", "", "", "")
results_queue.put((job_id, error_result))
def load_pdfs_async(file_links, model_choice, prompt_template, bm25_weight, temperature, top_p):
"""Asynchronous version of load_pdfs_updated to prevent timeouts"""
global last_job_id
if not file_links:
return "Please enter non-empty URLs", "", "Model used: N/A", "", "", get_job_list()
job_id = str(uuid.uuid4())
debug_print(f"Starting async job {job_id} for file loading")
# Start background thread
threading.Thread(
target=process_in_background,
args=(job_id, load_pdfs_updated, [file_links, model_choice, prompt_template, bm25_weight, temperature, top_p])
).start()
job_query = f"Loading files: {file_links.split()[0]}..." if file_links else "No files"
jobs[job_id] = {
"status": "processing",
"type": "load_files",
"start_time": time.time(),
"query": job_query
}
last_job_id = job_id
return (
f"Files submitted and processing in the background (Job ID: {job_id}).\n\n"
f"Use 'Check Job Status' tab with this ID to get results.",
f"Job ID: {job_id}",
f"Model requested: {model_choice}",
job_id, # Return job_id to update the job_id_input component
job_query, # Return job_query to update the job_query_display component
get_job_list() # Return updated job list
)
def submit_query_async(query, use_llama, use_mistral, temperature, top_p):
"""Asynchronous version of submit_query_updated to prevent timeouts"""
global last_job_id
if not query:
return ("Please enter a non-empty query", "Input/Output tokens: 0/0",
"Please enter a non-empty query", "Input/Output tokens: 0/0",
"", "", get_job_list())
if not (use_llama or use_mistral):
return ("Please select at least one model", "Input/Output tokens: 0/0",
"Please select at least one model", "Input/Output tokens: 0/0",
"", "", get_job_list())
responses = {"llama": None, "mistral": None}
job_ids = []
if use_llama:
llama_job_id = str(uuid.uuid4())
debug_print(f"Starting async job {llama_job_id} for Llama query: {query}")
# Start background thread for Llama
threading.Thread(
target=process_in_background,
args=(llama_job_id, submit_query_updated, [query, "🇺🇸 Remote Meta-Llama-3", temperature, top_p])
).start()
jobs[llama_job_id] = {
"status": "processing",
"type": "query",
"start_time": time.time(),
"query": query,
"model": "Llama"
}
job_ids.append(llama_job_id)
responses["llama"] = f"Processing (Job ID: {llama_job_id})"
if use_mistral:
mistral_job_id = str(uuid.uuid4())
debug_print(f"Starting async job {mistral_job_id} for Mistral query: {query}")
# Start background thread for Mistral
threading.Thread(
target=process_in_background,
args=(mistral_job_id, submit_query_updated, [query, "🇪🇺 Mistral-API", temperature, top_p])
).start()
jobs[mistral_job_id] = {
"status": "processing",
"type": "query",
"start_time": time.time(),
"query": query,
"model": "Mistral"
}
job_ids.append(mistral_job_id)
responses["mistral"] = f"Processing (Job ID: {mistral_job_id})"
# Store the last job ID (use the first one for now)
last_job_id = job_ids[0] if job_ids else None
return (
responses.get("llama", "Not selected"),
"Input tokens: " + str(count_tokens(query)) if use_llama else "Not selected",
responses.get("mistral", "Not selected"),
"Input tokens: " + str(count_tokens(query)) if use_mistral else "Not selected",
last_job_id,
query,
get_job_list()
)
def update_ui_with_last_job_id():
# This function doesn't need to do anything anymore
# We'll update the UI directly in the functions that call this
pass
# Function to display all jobs as a clickable list
def get_job_list():
job_list_md = "### Submitted Jobs\n\n"
if not jobs:
return "No jobs found. Submit a query or load files to create jobs."
# Sort jobs by start time (newest first)
sorted_jobs = sorted(
[(job_id, job_info) for job_id, job_info in jobs.items()],
key=lambda x: x[1].get("start_time", 0),
reverse=True
)
for job_id, job_info in sorted_jobs:
status = job_info.get("status", "unknown")
job_type = job_info.get("type", "unknown")
query = job_info.get("query", "")
model = job_info.get("model", "") # Get the model name
start_time = job_info.get("start_time", 0)
time_str = datetime.datetime.fromtimestamp(start_time).strftime("%Y-%m-%d %H:%M:%S")
# Create a shortened query preview
query_preview = query[:30] + "..." if query and len(query) > 30 else query or "N/A"
# Add color and icons based on status
if status == "processing":
status_formatted = f"<span style='color: red'>⏳ {status}</span>"
elif status == "completed":
status_formatted = f"<span style='color: green'>✅ {status}</span>"
else:
status_formatted = f"<span style='color: orange'>❓ {status}</span>"
# Add model icon based on model name
model_icon = "🇺🇸" if model == "Llama" else "🇪🇺" if model == "Mistral" else ""
model_prefix = f"{model_icon} {model} " if model else ""
# Create clickable links using Markdown
if job_type == "query":
job_list_md += f"- [{job_id}](javascript:void) - {time_str} - {status_formatted} - {model_prefix}Query: {query_preview}\n"
else:
job_list_md += f"- [{job_id}](javascript:void) - {time_str} - {status_formatted} - File Load Job\n"
return job_list_md
# Function to handle job list clicks
def job_selected(job_id):
if job_id in jobs:
return job_id, jobs[job_id].get("query", "No query for this job")
return job_id, "Job not found"
# Function to refresh the job list
def refresh_job_list():
return get_job_list()
# Function to sync model dropdown boxes
def sync_model_dropdown(value):
return value
# Function to check job status
def check_job_status(job_id):
if not job_id:
return "Please enter a job ID", "", "", "", ""
# Process any completed jobs in the queue
try:
while not results_queue.empty():
completed_id, result = results_queue.get_nowait()
if completed_id in jobs:
jobs[completed_id]["status"] = "completed"
jobs[completed_id]["result"] = result
jobs[completed_id]["end_time"] = time.time()
debug_print(f"Job {completed_id} completed and stored in jobs dictionary")
except queue.Empty:
pass
# Check if the requested job exists
if job_id not in jobs:
return "Job not found. Please check the ID and try again.", "", "", "", ""
job = jobs[job_id]
job_query = job.get("query", "No query available for this job")
# If job is still processing
if job["status"] == "processing":
elapsed_time = time.time() - job["start_time"]
job_type = job.get("type", "unknown")
if job_type == "load_files":
return (
f"Files are still being processed (elapsed: {elapsed_time:.1f}s).\n\n"
f"Try checking again in a few seconds.",
f"Job ID: {job_id}",
f"Status: Processing",
"",
job_query
)
else: # query job
return (
f"Query is still being processed (elapsed: {elapsed_time:.1f}s).\n\n"
f"Try checking again in a few seconds.",
f"Job ID: {job_id}",
f"Input tokens: {count_tokens(job.get('query', ''))}",
"Output tokens: pending",
job_query
)
# If job is completed
if job["status"] == "completed":
result = job["result"]
processing_time = job["end_time"] - job["start_time"]
if job.get("type") == "load_files":
return (
f"{result[0]}\n\nProcessing time: {processing_time:.1f}s",
result[1],
result[2],
"",
job_query
)
else: # query job
return (
f"{result[0]}\n\nProcessing time: {processing_time:.1f}s",
result[1],
result[2],
result[3],
job_query
)
# Fallback for unknown status
return f"Job status: {job['status']}", "", "", "", job_query
# Function to clean up old jobs
def cleanup_old_jobs():
current_time = time.time()
to_delete = []
for job_id, job in jobs.items():
# Keep completed jobs for 24 hours, processing jobs for 48 hours
if job["status"] == "completed" and (current_time - job.get("end_time", 0)) > 86400:
to_delete.append(job_id)
elif job["status"] == "processing" and (current_time - job.get("start_time", 0)) > 172800:
to_delete.append(job_id)
for job_id in to_delete:
del jobs[job_id]
debug_print(f"Cleaned up {len(to_delete)} old jobs. {len(jobs)} jobs remaining.")
return f"Cleaned up {len(to_delete)} old jobs", "", ""
# Improve the truncate_prompt function to be more aggressive with limiting context
def truncate_prompt(prompt: str, max_tokens: int = 4096) -> str:
"""Truncate prompt to fit within token limit, preserving the most recent/relevant parts."""
if not prompt:
return ""
if global_tokenizer:
try:
tokens = global_tokenizer.encode(prompt)
if len(tokens) > max_tokens:
# For prompts, we often want to keep the beginning instructions and the end context
# So we'll keep the first 20% and the last 80% of the max tokens
beginning_tokens = int(max_tokens * 0.2)
ending_tokens = max_tokens - beginning_tokens
new_tokens = tokens[:beginning_tokens] + tokens[-(ending_tokens):]
return global_tokenizer.decode(new_tokens)
except Exception as e:
debug_print(f"Truncation error: {str(e)}")
# Fallback to word-based truncation
words = prompt.split()
if len(words) > max_tokens:
beginning_words = int(max_tokens * 0.2)
ending_words = max_tokens - beginning_words
return " ".join(words[:beginning_words] + words[-(ending_words):])
return prompt
default_prompt = """\
{conversation_history}
Use the following context to provide a detailed technical answer to the user's question.
Do not include an introduction like "Based on the provided documents, ...". Just answer the question.
If you don't know the answer, please respond with "I don't know".
Context:
{context}
User's question:
{question}
"""
def load_txt_from_url(url: str) -> Document:
response = requests.get(url)
if response.status_code == 200:
text = response.text.strip()
if not text:
raise ValueError(f"TXT file at {url} is empty.")
return Document(page_content=text, metadata={"source": url})
else:
raise Exception(f"Failed to load {url} with status {response.status_code}")
class RemoteLLM(LLM):
temperature: float = 0.5
top_p: float = 0.95
def __init__(self, temperature: float = 0.5, top_p: float = 0.95):
super().__init__()
self.temperature = temperature
self.top_p = top_p
@property
def _llm_type(self) -> str:
return "remote_llm"
def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
try:
response = requests.post(
"http://localhost:11434/api/generate",
json={
"model": "llama2",
"prompt": prompt,
"temperature": self.temperature,
"top_p": self.top_p
},
stream=False
)
if response.status_code == 200:
return response.json()["response"]
else:
return f"Error: {response.status_code}"
except Exception as e:
return f"Error: {str(e)}"
@property
def _identifying_params(self) -> dict:
return {
"temperature": self.temperature,
"top_p": self.top_p
}
class MistralLLM(LLM):
temperature: float = 0.7
top_p: float = 0.95
_client: Any = PrivateAttr(default=None)
def __init__(self, api_key: str, temperature: float = 0.7, top_p: float = 0.95, **kwargs: Any):
try:
super().__init__(**kwargs)
object.__setattr__(self, '_client', Mistral(api_key=api_key))
self.temperature = temperature
self.top_p = top_p
except Exception as e:
debug_print(f"Init Mistral failed with error: {e}")
@property
def _llm_type(self) -> str:
return "mistral_llm"
def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
try:
debug_print("Calling Mistral API...")
response = self._client.chat.complete(
model="mistral-small-latest",
messages=[{"role": "user", "content": prompt}],
temperature=self.temperature,
top_p=self.top_p
)
return response.choices[0].message.content
except Exception as e:
debug_print(f"Mistral API error: {str(e)}")
return f"Error generating response: {str(e)}"
@property
def _identifying_params(self) -> dict:
return {"model": "mistral-small-latest"}
class LocalLLM(LLM):
@property
def _llm_type(self) -> str:
return "local_llm"
def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
truncated_prompt = truncate_prompt(prompt)
return f"Local LLM Fallback Response for: {truncated_prompt[:100]}..."
@property
def _identifying_params(self) -> dict:
return {}
class ErrorLLM(LLM):
@property
def _llm_type(self) -> str:
return "error_llm"
def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
return "Error: LLM pipeline could not be created. Please check your configuration and try again."
@property
def _identifying_params(self) -> dict:
return {}
class SimpleLLMChain:
def __init__(self, llm_choice: str = "Meta-Llama-3",
temperature: float = 0.5,
top_p: float = 0.95) -> None:
self.llm_choice = llm_choice
self.temperature = temperature
self.top_p = top_p
self.llm = self.create_llm_pipeline()
self.conversation_history = [] # Keep track of conversation
def create_llm_pipeline(self):
from langchain.llms.base import LLM # Import LLM here so it's always defined
normalized = self.llm_choice.lower()
try:
if "remote" in normalized:
debug_print("Creating remote Meta-Llama-3 pipeline via Hugging Face Inference API...")
from huggingface_hub import InferenceClient
repo_id = "meta-llama/Meta-Llama-3-8B-Instruct"
hf_api_token = os.environ.get("HF_API_TOKEN")
if not hf_api_token:
raise ValueError("Please set the HF_API_TOKEN environment variable to use remote inference.")
client = InferenceClient(token=hf_api_token, timeout=120)
# We no longer use wait_for_model because it's unsupported
def remote_generate(prompt: str) -> str:
max_retries = 3
backoff = 2 # start with 2 seconds
for attempt in range(max_retries):
try:
debug_print(f"Remote generation attempt {attempt+1}")
response = client.text_generation(
prompt,
model=repo_id,
temperature=self.temperature,
top_p=self.top_p,
max_new_tokens=512 # Reduced token count for speed
)
return response
except Exception as e:
debug_print(f"Attempt {attempt+1} failed with error: {e}")
if attempt == max_retries - 1:
raise
time.sleep(backoff)
backoff *= 2 # exponential backoff
return "Failed to generate response after multiple attempts."
class RemoteLLM(LLM):
@property
def _llm_type(self) -> str:
return "remote_llm"
def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
return remote_generate(prompt)
@property
def _identifying_params(self) -> dict:
return {"model": repo_id}
debug_print("Remote Meta-Llama-3 pipeline created successfully.")
return RemoteLLM()
elif "mistral" in normalized:
api_key = os.getenv("MISTRAL_API_KEY")
return MistralLLM(api_key=api_key, temperature=self.temperature, top_p=self.top_p)
else:
return LocalLLM()
except Exception as e:
debug_print(f"Error creating LLM pipeline: {str(e)}")
return ErrorLLM()
def update_llm_pipeline(self, new_model_choice: str, temperature: float, top_p: float):
self.llm_choice = new_model_choice
self.temperature = temperature
self.top_p = top_p
self.llm = self.create_llm_pipeline()
def submit_query(self, query: str) -> tuple:
try:
response = self.llm(query)
# Store in conversation history
self.conversation_history.append({"query": query, "response": response})
input_tokens = count_tokens(query)
output_tokens = count_tokens(response)
return (response, f"Input tokens: {input_tokens}", f"Output tokens: {output_tokens}")
except Exception as e:
return (f"Error processing query: {str(e)}", "Input tokens: 0", "Output tokens: 0")
# Update submit_query_updated to work with the simplified chain
def submit_query_updated(query: str, model_choice: str = None, temperature: float = 0.5, top_p: float = 0.95):
"""Process a query with the specified model and parameters."""
debug_print(f"Processing query: {query}")
if not query:
debug_print("Empty query received")
return "Please enter a non-empty query", "", "Input tokens: 0", "Output tokens: 0"
try:
global llm_chain
if llm_chain is None:
llm_chain = SimpleLLMChain(
llm_choice=model_choice,
temperature=temperature,
top_p=top_p
)
elif llm_chain.llm_choice != model_choice:
llm_chain.update_llm_pipeline(model_choice, temperature, top_p)
response, input_tokens, output_tokens = llm_chain.submit_query(query)
return response, "", input_tokens, output_tokens
except Exception as e:
debug_print(f"Error in submit_query_updated: {str(e)}")
return f"Error: {str(e)}", "", "Input tokens: 0", "Output tokens: 0"
# Update the reset_app_updated function
def reset_app_updated():
global llm_chain
llm_chain = None
return "Application reset successfully"
# ----------------------------
# Gradio Interface Functions
# ----------------------------
global rag_chain
rag_chain = SimpleLLMChain()
def load_pdfs_updated(file_links, model_choice, prompt_template, bm25_weight, temperature, top_p):
debug_print("Inside load_pdfs function.")
if not file_links:
debug_print("Please enter non-empty URLs")
return "Please enter non-empty URLs", "Word count: N/A", "Model used: N/A", "Context: N/A"
try:
links = [link.strip() for link in file_links.split("\n") if link.strip()]
global rag_chain
if rag_chain.raw_data:
rag_chain.update_llm_pipeline(model_choice, temperature, top_p, prompt_template, bm25_weight)
context_display = rag_chain.get_current_context()
response_msg = f"Files already loaded. Chain updated with model: {model_choice}"
return (
response_msg,
f"Word count: {word_count(rag_chain.context)}",
f"Model used: {rag_chain.llm_choice}",
f"Context:\n{context_display}"
)
else:
rag_chain = SimpleLLMChain(
llm_choice=model_choice,
temperature=temperature,
top_p=top_p
)
rag_chain.add_pdfs_to_vectore_store(links)
context_display = rag_chain.get_current_context()
response_msg = f"Files loaded successfully. Using model: {model_choice}"
return (
response_msg,
f"Word count: {word_count(rag_chain.context)}",
f"Model used: {rag_chain.llm_choice}",
f"Context:\n{context_display}"
)
except Exception as e:
error_msg = traceback.format_exc()
debug_print("Could not load files. Error: " + error_msg)
return (
"Error loading files: " + str(e),
f"Word count: {word_count('')}",
f"Model used: {rag_chain.llm_choice}",
"Context: N/A"
)
def update_model(new_model: str):
global rag_chain
if rag_chain and rag_chain.raw_data:
rag_chain.update_llm_pipeline(new_model, rag_chain.temperature, rag_chain.top_p,
rag_chain.prompt_template, rag_chain.bm25_weight)
debug_print(f"Model updated to {rag_chain.llm_choice}")
return f"Model updated to: {rag_chain.llm_choice}"
else:
return "No files loaded; please load files first."
def reset_app_updated():
global rag_chain
rag_chain = SimpleLLMChain()
debug_print("App reset successfully.")
return (
"App reset successfully. You can now load new files",
"",
"Model used: Not selected"
)
# ----------------------------
# Gradio Interface Setup
# ----------------------------
custom_css = """
textarea {
overflow-y: scroll !important;
max-height: 200px;
}
"""
# Function to add dots and reset
def add_dots_and_reset():
if not hasattr(add_dots_and_reset, "dots"):
add_dots_and_reset.dots = "" # Initialize the attribute
# Add a dot
add_dots_and_reset.dots += "."
# Reset after 5 dots
if len(add_dots_and_reset.dots) > 5:
add_dots_and_reset.dots = ""
print(f"Current dots: {add_dots_and_reset.dots}") # Debugging print
return add_dots_and_reset.dots
# Define a dummy function to simulate data retrieval
def run_query(max_value):
# Simulate a data retrieval or processing function
return [[i, i**2] for i in range(1, max_value + 1)]
# Function to call both refresh_job_list and check_job_status using the last job ID
def periodic_update(is_checked):
interval = 2 if is_checked else None
debug_print(f"Auto-refresh checkbox is {'checked' if is_checked else 'unchecked'}, every={interval}")
if is_checked:
global last_job_id
job_list_md = refresh_job_list()
job_status = check_job_status(last_job_id) if last_job_id else ("No job ID available", "", "", "", "")
query_results = run_query(10) # Use a fixed value or another logic if needed
return job_list_md, job_status[0], query_results, "" # Return empty string instead of context
else:
# Return empty values to stop updates
return "", "", [], ""
# Define a function to determine the interval based on the checkbox state
def get_interval(is_checked):
return 2 if is_checked else None
# Update the Gradio interface to include job status checking
with gr.Blocks(css=custom_css, js="""
document.addEventListener('DOMContentLoaded', function() {
// Add event listener for job list clicks
const jobListInterval = setInterval(() => {
const jobLinks = document.querySelectorAll('.job-list-container a');
if (jobLinks.length > 0) {
jobLinks.forEach(link => {
link.addEventListener('click', function(e) {
e.preventDefault();
const jobId = this.textContent.split(' ')[0];
// Find the job ID input textbox and set its value
const jobIdInput = document.querySelector('.job-id-input input');
if (jobIdInput) {
jobIdInput.value = jobId;
// Trigger the input event to update Gradio's state
jobIdInput.dispatchEvent(new Event('input', { bubbles: true }));
}
});
});
clearInterval(jobListInterval);
}
}, 500);
});
""") as app:
gr.Markdown('''# PsyLLM Interface
**Model Selection & Parameters:** Choose from the following options:
- 🇺🇸 Remote Meta-Llama-3 - has context windows of 8000 tokens
- 🇪🇺 Mistral-API - has context windows of 32000 tokens
**🔥 Randomness (Temperature):** Adjusts output predictability.
- Example: 0.2 makes the output very deterministic (less creative), while 0.8 introduces more variety and spontaneity.
**🎯 Word Variety (Top‑p):** Limits word choices to a set probability percentage.
- Example: 0.5 restricts output to the most likely 50% of token choices for a focused answer; 0.95 allows almost all possibilities for more diverse responses.
**⚠️ IMPORTANT: This app uses asynchronous processing to avoid timeout issues**
- When you submit a query, you'll receive a Job ID
- Use the "Check Job Status" tab to monitor and retrieve your results
''')
with gr.Tabs() as tabs:
with gr.TabItem("Submit Query"):
with gr.Row():
with gr.Column(scale=1):
llama_checkbox = gr.Checkbox(
value=True,
label="🇺🇸 Remote Meta-Llama-3",
info="Context window: 8000 tokens"
)
mistral_checkbox = gr.Checkbox(
value=False,
label="🇪🇺 Mistral-API",
info="Context window: 32000 tokens"
)
with gr.Column(scale=2):
temperature_slider = gr.Slider(
minimum=0.1, maximum=1.0, value=0.5, step=0.1,
label="Randomness (Temperature)"
)
top_p_slider = gr.Slider(
minimum=0.1, maximum=0.99, value=0.95, step=0.05,
label="Word Variety (Top-p)"
)
with gr.Row():
query_input = gr.Textbox(
label="Enter your query here",
placeholder="Type your query",
lines=4
)
submit_button = gr.Button("Submit Query to Selected Models")
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("### Llama Results")
llama_response = gr.Textbox(
label="Llama Response",
placeholder="Response will appear here",
lines=6
)
llama_tokens = gr.Markdown("Input/Output tokens: 0/0")
with gr.Column(scale=1):
gr.Markdown("### Mistral Results")
mistral_response = gr.Textbox(
label="Mistral Response",
placeholder="Response will appear here",
lines=6
)
mistral_tokens = gr.Markdown("Input/Output tokens: 0/0")
with gr.TabItem("Check Job Status"):
with gr.Row():
with gr.Column(scale=1):
job_list = gr.Markdown(
value="No jobs yet",
label="Job List (Click to select)"
)
# Add the Refresh Job List button
refresh_button = gr.Button("Refresh Job List")
# Use a Checkbox to control the periodic updates
auto_refresh_checkbox = gr.Checkbox(
label="Enable Auto Refresh",
value=False # Default to unchecked
)
# Use a DataFrame to display results
df = gr.DataFrame(
value=run_query(10), # Initial value
headers=["Number", "Square"],
label="Query Results",
visible=False # Set the DataFrame to be invisible
)
with gr.Column(scale=2):
job_id_input = gr.Textbox(
label="Job ID",
placeholder="Job ID will appear here when selected from the list",
lines=1
)
job_query_display = gr.Textbox(
label="Job Query",
placeholder="The query associated with this job will appear here",
lines=2,
interactive=False
)
check_button = gr.Button("Check Status")
cleanup_button = gr.Button("Cleanup Old Jobs")
with gr.Row():
status_response = gr.Textbox(
label="Job Result",
placeholder="Job result will appear here",
lines=6
)
status_context = gr.Textbox(
label="Context Information",
placeholder="Context information will appear here",
lines=6
)
with gr.Row():
status_tokens1 = gr.Markdown("")
status_tokens2 = gr.Markdown("")
with gr.TabItem("App Management"):
with gr.Row():
reset_button = gr.Button("Reset App")
with gr.Row():
reset_response = gr.Textbox(
label="Reset Response",
placeholder="Reset confirmation will appear here",
lines=2
)
reset_context = gr.Textbox(
label="",
placeholder="",
lines=2,
visible=False
)
with gr.Row():
reset_model = gr.Markdown("")
# Connect the buttons to their respective functions
submit_button.click(
submit_query_async,
inputs=[
query_input,
llama_checkbox,
mistral_checkbox,
temperature_slider,
top_p_slider
],
outputs=[
llama_response,
llama_tokens,
mistral_response,
mistral_tokens,
job_id_input,
job_query_display,
job_list
]
)
check_button.click(
check_job_status,
inputs=[job_id_input],
outputs=[status_response, status_context, status_tokens1, status_tokens2, job_query_display]
)
refresh_button.click(
refresh_job_list,
inputs=[],
outputs=[job_list]
)
job_id_input.change(
job_selected,
inputs=[job_id_input],
outputs=[job_id_input, job_query_display]
)
cleanup_button.click(
cleanup_old_jobs,
inputs=[],
outputs=[status_response, status_context, status_tokens1]
)
reset_button.click(
reset_app_updated,
inputs=[],
outputs=[reset_response, reset_context, reset_model]
)
app.load(
fn=refresh_job_list,
inputs=None,
outputs=job_list
)
auto_refresh_checkbox.change(
fn=periodic_update,
inputs=[auto_refresh_checkbox],
outputs=[job_list, status_response, df, status_context],
every=2
)
# Add this with your other global variables
global llm_chain
llm_chain = None
if __name__ == "__main__":
debug_print("Launching Gradio interface.")
app.queue().launch(share=False)