import gradio as gr import requests import os import json import time import transformers import re from transformers import AutoTokenizer, AutoModelForCausalLM hf_token = os.getenv("HF_AUTH_TOKEN") vapi_url = "https://api-inference.huggingface.co/models/vectara/hallucination_evaluation_model" headers = {"Authorization": f"Bearer {hf_token}"} model_name = "allenai/OLMo-1B" tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True) def generate_text(prompt, max_new_tokens=100, do_sample=False, top_k=50, top_p=0.95): inputs = tokenizer(prompt, return_tensors='pt', return_token_type_ids=False) response = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=do_sample, top_k=top_k, top_p=top_p) return tokenizer.batch_decode(response, skip_special_tokens=True)[0] # Function to query the API def query(payload): response = requests.post(vapi_url, headers=headers, json=payload) return response.json() def check_hallucination(assertion, citation): api_url = "https://api-inference.huggingface.co/models/vectara/hallucination_evaluation_model" header = {"Authorization": f"Bearer {hf_token}"} payload = {"inputs": f"{assertion} [SEP] {citation}"} attempts = 0 max_attempts = 3 wait_time = 180 # 3 minutes while attempts < max_attempts: try: response = requests.post(api_url, headers=header, json=payload, timeout=120) response.raise_for_status() # This will raise an exception for HTTP error codes output = response.json() output = output[0][0]["score"] return f"**hallucination score:** {output}" except requests.exceptions.HTTPError as http_err: print(f"HTTP error occurred: {http_err}") # Python 3.6 except requests.exceptions.RequestException as err: print(f"Other error occurred: {err}") # Python 3.6 except KeyError: print("KeyError: The expected key was not found in the response. The endpoint might be waking up.") attempts += 1 if attempts < max_attempts: print(f"Attempt {attempts} failed. Waiting for {wait_time} seconds before retrying...") time.sleep(wait_time) else: print("Maximum attempts reached. Please try again later.") return "Error: Unable to retrieve hallucination score after multiple attempts." return "Error: Unable to process the hallucination check." def query_vectara(text): user_message = text customer_id = os.getenv('CUSTOMER_ID') corpus_id = os.getenv('CORPUS_ID') api_key = os.getenv('API_KEY') api_key_header = { "customer-id": customer_id, "x-api-key": api_key } request_body = { "query": [ { "query": user_message, "queryContext": "", "start": 1, "numResults": 25, "contextConfig": { "charsBefore": 0, "charsAfter": 0, "sentencesBefore": 2, "sentencesAfter": 2, "startTag": "%START_SNIPPET%", "endTag": "%END_SNIPPET%", }, "rerankingConfig": { "rerankerId": 272725718, "mmrConfig": { "diversityBias": 0.35 } }, "corpusKey": [ { "customerId": customer_id, "corpusId": corpus_id, "semantics": 0, "metadataFilter": "", "lexicalInterpolationConfig": { "lambda": 0 }, "dim": [] } ], "summary": [ { "maxSummarizedResults": 5, "responseLang": "auto", "summarizerPromptName": "vectara-summary-ext-v1.2.0" } ] } ] } response = requests.post( "https://api.vectara.io/v1/query", json=request_body, verify=True, headers=api_key_header ) if response.status_code == 200: query_data = response.json() if query_data: sources_info = [] # Extract the summary. summary = query_data['responseSet'][0]['summary'][0]['text'] # Iterate over all response sets for response_set in query_data.get('responseSet', []): # Extract sources # Limit to top 5 sources. for source in response_set.get('response', [])[:5]: source_metadata = source.get('metadata', []) source_info = {} for metadata in source_metadata: metadata_name = metadata.get('name', '') metadata_value = metadata.get('value', '') if metadata_name == 'title': source_info['title'] = metadata_value elif metadata_name == 'author': source_info['author'] = metadata_value elif metadata_name == 'pageNumber': source_info['page number'] = metadata_value if source_info: sources_info.append(source_info) result = {"summary": summary, "sources": sources_info} return f"{json.dumps(result, indent=2)}" else: return "No data found in the response." else: return f"Error: {response.status_code}" def remove_references(text): # Regex pattern to find references like [1], [1][2], etc. pattern = r'\[\d+\]+' # Replace found patterns with an empty string cleaned_text = re.sub(pattern, '', text) return cleaned_text def clean_text(text): # Remove special characters, keeping only letters, numbers, and spaces cleaned_text = re.sub(r'[^a-zA-Z0-9\s]', '', text) return cleaned_text def evaluate_content(user_input): vectara_response = query_vectara(user_input) vectara_response_json = json.loads(vectara_response) summary = vectara_response_json.get("summary", "") sources = vectara_response_json.get("sources", []) # Remove references from the summary text summary_no_refs = remove_references(summary) # Clean summary text to remove special characters summary_clean = clean_text(summary_no_refs) # Process sources to extract and clean necessary information sources_info = "" for source in sources: title = source.get("title", "No title") author = source.get("author", "No author") page_number = source.get("page number", "N/A") # Clean source info title_clean = clean_text(title) author_clean = clean_text(author) sources_info += f"Title: {title_clean}, Author: {author_clean}, Page: {page_number}\n" # Generate text based on the cleaned and reference-removed summary olmo_output = generate_text(summary_clean) olmo_output_clean = clean_text(olmo_output) # Check hallucination based on the original output and summary hallucination_score = check_hallucination(olmo_output, summary) return summary_clean, sources_info, olmo_output_clean, hallucination_score # Adjust the Gradio interface outputs to match the new structure iface = gr.Interface( fn=evaluate_content, inputs=[gr.Textbox(label="User Input")], outputs=[ gr.Textbox(label="Vectara Summary", lines=10), gr.Textbox(label="Vectara Sources", lines=10), gr.Textbox(label="Generated Text", lines=10), gr.Textbox(label="Hallucination Score") ], live=False, title="👋🏻Welcome to 🌟Team Tonic's 🧠🌈SureRAG🔴🟢", description="Nothing is more important than reputation. However you can create automated content pipelines for public facing content. How can businesses grow their reputation while mitigating risks due to AI? How it works : vectara rag retrieval reranking and summarization is used to return content. then an LLM generates content based on these returns. this content is checked for hallucination before being validated for publishing on twitter. SureRAG is fixed on Tonic-AI's README files as a Demo, provide input to generate a response. This response is checked by Vectara's HHME. Check out the model [vectara/hallucination_evaluation_model](https://huggingface.co/vectara/hallucination_evaluation_model) Join us : 🌟TeamTonic🌟 is always making cool demos! Join our active builder's🛠️community 👻 [![Join us on Discord](https://img.shields.io/discord/1109943800132010065?label=Discord&logo=discord&style=flat-square)](https://discord.gg/GWpVpekp) On 🤗Huggingface: [TeamTonic](https://huggingface.co/TeamTonic) & [MultiTransformer](https://huggingface.co/MultiTransformer) On 🌐Github: [Tonic-AI](https://github.com/tonic-ai) & contribute to 🌟 [DataTonic](https://github.com/Tonic-AI/DataTonic)", ) iface.launch()