from datasets import load_from_disk from transformers import AutoTokenizer, AutoModel import faiss import numpy as np import torch from datasets import load_from_disk import faiss import numpy as np import os from datasets import load_dataset, Dataset, get_dataset_config_names from sentence_transformers import SentenceTransformer from groq import Groq from sentence_transformers import CrossEncoder import requests import uuid import re import json import gradio as gr import io import sys import traceback embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2") # Preload datasets and indices hf_dataset_cs = load_from_disk("cs_dataset") faiss_index_cs = faiss.read_index("cs_index/faiss.index") hf_dataset_med = load_from_disk("med_dataset") faiss_index_med = faiss.read_index("med_index/faiss.index") hf_dataset_gk = load_from_disk("gk_dataset") faiss_index_gk = faiss.read_index("gk_index/faiss.index") hf_dataset_fin = load_from_disk("fin_dataset") faiss_index_fin = faiss.read_index("fin_index/faiss.index") device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(device) legal_dataset = load_dataset("rungalileo/ragbench", "cuad", split="test") med_dataset = load_dataset("rungalileo/ragbench", "pubmedqa", split="test") gk_dataset = load_dataset("rungalileo/ragbench", "hotpotqa", split="test") cs_dataset = load_dataset("rungalileo/ragbench", "emanual", split="test") fin_dataset = load_dataset("rungalileo/ragbench", "finqa", split="test") # Load BGE reranker reranker = CrossEncoder("BAAI/bge-reranker-base", max_length=512) embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2") model_name = "nlpaueb/legal-bert-base-uncased" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModel.from_pretrained(model_name).to(device) model.eval() def retrieve_top_k(query,domain='legal', model_name='nlpaueb/legal-bert-base-uncased', k=8): # Load tokenizer and model tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModel.from_pretrained(model_name).to(device) model.eval() #print(f"In retrive_top_k Query:{query}") # Tokenize and embed query using mean pooling inputs = tokenizer(query, return_tensors="pt", padding=True, truncation=True, max_length=512) inputs = {k: v.to(device) for k, v in inputs.items()} with torch.no_grad(): outputs = model(**inputs) query_embedding = outputs.last_hidden_state.mean(dim=1).cpu().numpy() # Load FAISS index and dataset index_path = f"legal_index/faiss.index" dataset_path = f"legal_dataset" faiss_index = faiss.read_index(index_path) dataset = load_from_disk(dataset_path) # Perform FAISS search D, I = faiss_index.search(query_embedding.astype('float32'), k) # Retrieve top-k matching chunks top_chunks = [dataset[int(idx)]['text'] for idx in I[0]] return top_chunks # Retrieval function using preloaded objects def retrieve_top_c(query, domain, embedder, k=5): if domain == "CS": hf_dataset = hf_dataset_cs faiss_index = faiss_index_cs elif domain == "Medical": hf_dataset = hf_dataset_med faiss_index = faiss_index_med elif domain == "GK": hf_dataset = hf_dataset_gk faiss_index = faiss_index_gk elif domain == "Finance": hf_dataset = hf_dataset_fin faiss_index = faiss_index_fin else: raise ValueError(f"Unknown domain: {domain}") # Encode query and search query_embedding = embedder.encode([query]).astype('float32') #query_embedding = embedder.encode([query], convert_to_numpy=True).astype('float32') distances, indices = faiss_index.search(query_embedding, k) return [hf_dataset[int(i)]["text"] for i in indices[0]] client = Groq( api_key= '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',#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', ) def rerank_documents_bge(query, documents, top_n=5, return_scores=False): """ Rerank documents using BAAI/bge-reranker-base CrossEncoder. Args: query (str): The query string. documents (List[str]): List of candidate documents. top_n (int): Number of top results to return. return_scores (bool): Whether to return scores along with documents. Returns: List[str] or List[Tuple[str, float]] """ if not documents: return [] # Prepare (query, doc) pairs pairs = [(query, doc) for doc in documents] # Predict relevance scores scores = reranker.predict(pairs, batch_size=16) # Sort by score descending reranked = sorted(zip(documents, scores), key=lambda x: x[1], reverse=True) if return_scores: return reranked[:top_n] else: return [doc for doc, _ in reranked[:top_n]] def generate_response_rag(query,domain): # Step 1: Retrieve top-k context chunks using your FAISS setup if domain == "Legal": top_chunks = retrieve_top_k(query,'Legal', model_name) else: top_chunks = retrieve_top_c(query, domain,embedder) # Step 2: Rerank retrieved documents using cross-encoder #reranked_chunks = rerank_documents(query, top_chunks, top_n=15) #rerank_and_filter_chunks = filter_by_faithfulness(query, reranked_chunks) #print("Retrieved Top chunks",top_chunks) #reranked_chunks = rerank_and_filter_chunks reranked_chunks_bge = rerank_documents_bge(query, top_chunks, top_n=5) #sum_context = summarize_context("\n\n".join(reranked_chunks_bge)) final_context = reranked_chunks_bge # Step 2: Prepare context and RAG-style prompt context = "\n\n".join(final_context) #print(f"Context:{context}") prompt = f"""You are a helpful legal assistant. Use the following context to answer the question. Using only the information from the retrieved context, answer the following question. If the answer cannot be derived, say "I don't know." Always have answer with prefix **Answer:** Context:{context} Question: {query} Answer:""" # Step 3: Call the LLM (LLaMA3 or any chat model) chat_completion = client.chat.completions.create( messages=[ {"role": "user", "content": prompt} ], model="llama3-70b-8192",#"gemma2-9b-it"#"qwen/qwen3-32b"#deepseek-r1-distill-llama-70b",#"llama3-70b-8192", # mistral-saba-24b temperature=0.0 ) return context,chat_completion.choices[0].message.content.strip() '''response = openai.chat.completions.create( model="gpt-3.5-turbo", messages=[ {"role": "user", "content": prompt} ], temperature=0.0, max_tokens=1024 ) return response.choices[0].message.content''' #JUDGE LLM def split_into_keyed_sentences(text, prefix): """Splits text into sentences with keys like '0a.', '0b.', or 'a.', 'b.', etc.""" # Basic sentence tokenizer with keys sentences = re.split(r'(?<=[.?!])\s+', text.strip()) keyed = {} for i, s in enumerate(sentences): key = f"{prefix}{chr(97 + i)}" # 'a', 'b', ... if s: keyed[key] = s.strip() return keyed def jugde_response_rag(query, domain): #top_chunks = retrieve_top_k(query) #top_chunks = [chunk[0] if isinstance(chunk, tuple) else chunk for chunk in top_chunks] # Step 2: Prepare context and RAG-style prompt #context = "\n\n".join(top_chunks) # Split context and dummy answer into keyed sentences #document_keys = split_into_keyed_sentences(context, "0") #print(f"Query:{query}\n====================================================================") context,response = generate_response_rag(query,domain) #deepseek-r1-distill-llama-70b llama3-70b-8192 # Split context and dummy answer into keyed sentences document_keys = split_into_keyed_sentences(context, "0") #print(f"\n====================================\Generator Response:{response}") #For deepseek #print("Before Curated:",response) response=response[response.find("**Answer"):].replace("**Answer",""); print(f"Response for Generator LLM:{response}") response_keys = split_into_keyed_sentences(response, "") # Rebuild sections for prompt documents_formatted = "\n".join([f"{k}. {v}" for k, v in document_keys.items()]) response_formatted = "\n".join([f"{k}. {v}" for k, v in response_keys.items()]) '''print(f"\n====================================================================") print(f"documents_formatted:{documents_formatted}") print(f"\n====================================================================") print(f"response_formatted:{response_formatted}") print(f"\n====================================================================")''' prompt = f"""I asked someone to answer a question based on one or more documents. Your task is to review their response and assess whether or not each sentence in that response is supported by text in the documents. And if so, which sentences in the documents provide that support. You will also tell me which of the documents contain useful information for answering the question, and which of the documents the answer was sourced from. Here are the documents, each of which is split into sentences. Alongside each sentence is associated key, such as ’0a.’ or ’0b.’ that you can use to refer to it: ''' {documents_formatted} ''' The question was: ''' {query} ''' Here is their response, split into sentences. Alongside each sentence is associated key, such as ’a.’ or ’b.’ that you can use to refer to it. Note that these keys are unique to the response, and are not related to the keys in the documents: ''' {response_formatted} ''' You must respond with a JSON object matching this schema: ''' {{ "relevance_explanation": string, "all_relevant_sentence_keys": [string], "overall_supported_explanation": string, "overall_supported": boolean, "sentence_support_information": [ {{ "response_sentence_key": string, "explanation": string, "supporting_sentence_keys": [string], "fully_supported": boolean }}, ], "all_utilized_sentence_keys": [string] }} ''' The relevance_explanation field is a string explaining which documents contain useful information for answering the question. Provide a step-by-step breakdown of information provided in the documents and how it is useful for answering the question. The all_relevant_sentence_keys field is a list of all document sentences keys (e.g. ’0a’) that are revant to the question. Include every sentence that is useful and relevant to the question, even if it was not used in the response, or if only parts of the sentence are useful. Ignore the provided response when making this judgement and base your judgement solely on the provided documents and question. Omit sentences that, if removed from the document, would not impact someone’s ability to answer the question. The overall_supported_explanation field is a string explaining why the response *as a whole* is or is not supported by the documents. In this field, provide a step-by-step breakdown of the claims made in the response and the support (or lack thereof) for those claims in the documents. Begin by assessing each claim separately, one by one; don’t make any remarks about the response as a whole until you have assessed all the claims in isolation. The overall_supported field is a boolean indicating whether the response as a whole is supported by the documents. This value should reflect the conclusion you drew at the end of your step-by-step breakdown in overall_supported_explanation. In the sentence_support_information field, provide information about the support *for each sentence* in the response. The sentence_support_information field is a list of objects, one for each sentence in the response. Each object MUST have the following fields: - response_sentence_key: a string identifying the sentence in the response. This key is the same as the one used in the response above. - explanation: a string explaining why the sentence is or is not supported by the documents. - supporting_sentence_keys: keys (e.g. ’0a’) of sentences from the documents that support the response sentence. If the sentence is not supported, this list MUST be empty. If the sentence is supported, this list MUST contain one or more keys. In special cases where the sentence is supported, but not by any specific sentence, you can use the string "supported_without_sentence" to indicate that the sentence is generally supported by the documents. Consider cases where the sentence is expressing inability to answer the question due to lack of relevant information in the provided contex as "supported_without_sentence". In cases where the sentence is making a general statement (e.g. outlining the steps to produce an answer, or summarizing previously stated sentences, or a transition sentence), use the sting "general".In cases where the sentence is correctly stating a well-known fact, like a mathematical formula, use the string "well_known_fact". In cases where the sentence is performing numerical reasoning (e.g. addition, multiplication), use the string "numerical_reasoning". - fully_supported: a boolean indicating whether the sentence is fully supported by the documents. - This value should reflect the conclusion you drew at the end of your step-by-step breakdown in explanation. - If supporting_sentence_keys is an empty list, then fully_supported must be false. 17 - Otherwise, use fully_supported to clarify whether everything in the response sentence is fully supported by the document text indicated in supporting_sentence_keys (fully_supported = true), or whether the sentence is only partially or incompletely supported by that document text (fully_supported = false). The all_utilized_sentence_keys field is a list of all sentences keys (e.g. ’0a’) that were used to construct the answer. Include every sentence that either directly supported the answer, or was implicitly used to construct the answer, even if it was not used in its entirety. Omit sentences that were not used, and could have been removed from the documents without affecting the answer. You must respond with a valid JSON string. Use escapes for quotes, e.g. ‘\\"‘, and newlines, e.g. ‘\\n‘. Do not write anything before or after the JSON string. Do not wrap the JSON string in backticks like ‘‘‘ or ‘‘‘json. As a reminder: your task is to review the response and assess which documents contain useful information pertaining to the question, and how each sentence in the response is supported by the text in the documents.\ """ # Step 3: Call the LLM chat_completion = client.chat.completions.create( messages=[ {"role": "user", "content": prompt} ], model="meta-llama/llama-4-maverick-17b-128e-instruct", #deepseek-r1-distill-llama-70b llama3-70b-8192 meta-llama/llama-4-maverick-17b-128e-instruct ) return documents_formatted,chat_completion.choices[0].message.content.strip() '''chat_completion = openai.chat.completions.create( messages=[ {"role":"user", "content":prompt} ], model="gpt-4o", max_tokens=1024, ) return documents_formatted,chat_completion.choices[0].message.content''' def extract_retrieved_sentence_keys(document_text: str) -> list[str]: """ Extracts sentence keys like '0a.', '0b.', etc. from a formatted document string. Parameters: - document_text (str): full text of document with sentence keys Returns: - List of unique sentence keys in the order they appear """ # Match pattern like 0a., 0b., 0z., 0{., 0|., etc. pattern = r'\b0[\w\{\|\}~€‚]\.' matches = re.findall(pattern, document_text) return list(dict.fromkeys(matches)) # Removes duplicates while preserving order def compute_ragbench_metrics(judge_response: dict, retrieved_sentence_keys: list[str]) -> dict: """ Computes RAGBench-style metrics from Judge LLM response. Parameters: - judge_response (dict): JSON response from Judge LLM - retrieved_sentence_keys (list of str): all sentence keys from the retrieved documents Returns: - Dictionary with Context Relevance, Context Utilization, Completeness, and Adherence """ R = set(judge_response.get("all_relevant_sentence_keys", [])) # Relevant sentences U = set(judge_response.get("all_utilized_sentence_keys", [])) # Utilized sentences intersection_RU = R & U total_retrieved = len(retrieved_sentence_keys) len_R = len(R) len_U = len(U) len_intersection = len(intersection_RU) # Context Relevance: fraction of retrieved context that is relevant context_relevance = len_R / total_retrieved if total_retrieved else 0.0 # Context Utilization: fraction of retrieved context that was used context_utilization = len_U / total_retrieved if total_retrieved else 0.0 # Completeness: fraction of relevant content that was used completeness = len_intersection / len_R if len_R else 0.0 # Adherence: 1 if all response sentences are fully supported, else 0 is_fully_supported = all(s.get("fully_supported", False) for s in judge_response.get("sentence_support_information", [])) adherence = 1.0 if is_fully_supported and judge_response.get("overall_supported", False) else 0.0 return { "Context Relevance": round(context_relevance, 4), "Context Utilization": round(context_utilization, 4), "Completeness": round(completeness, 4), "Adherence": adherence } def evaluate_rag_pipeline(domain, q_indices): import torch import numpy as np from sklearn.metrics import mean_squared_error, roc_auc_score device = torch.device("cuda" if torch.cuda.is_available() else "cpu") def safe_append(gt_list, pred_list, gt_val, pred_val): if gt_val is not None and pred_val is not None: gt_list.append(gt_val) pred_list.append(pred_val) def clean_and_parse_json_block(text): # Strip markdown-style code block if present #text = text.strip().strip("`").strip() code_block_match = re.search(r"```(?:json)?\s*([\s\S]*?)\s*```", text) if code_block_match: text = code_block_match.group(1).strip() # Remove invalid/control characters that break decoding text = re.sub(r"[^\x20-\x7E\n\t]", "", text) try: return json.loads(text) except json.JSONDecodeError as e: print("❌ JSON Decode Error:", e) print("⚠️ Cleaned text:\n", text) raise gt_relevance, pred_relevance = [], [] gt_utilization, pred_utilization = [], [] gt_completeness, pred_completeness = [], [] gt_adherence, pred_adherence = [], [] if(domain=="Legal"): dataset = legal_dataset elif(domain=="Medical"): dataset = med_dataset elif(domain=="GK"): dataset = gk_dataset elif(domain=="CS"): dataset = cs_dataset elif(domain=="Finance"): dataset = fin_dataset for i in q_indices: query = dataset[i]['question'] print(f"\n\n\nQuery:{i}.{query}\n====================================================================") #print(f"\ndomain:{domain}====================================================================") documents_formatted, response = jugde_response_rag(query, domain) judge_response = clean_and_parse_json_block(response) print(f"\ndocuments_formatted:{documents_formatted}") print(f"\n======================================================================\nResponse:{judge_response}") retrieved_sentences = extract_retrieved_sentence_keys(documents_formatted) predicted = compute_ragbench_metrics(judge_response, retrieved_sentences) # GT values gt_r = dataset[i].get('relevance_score') gt_u = dataset[i].get('utilization_score') gt_c = dataset[i].get('completeness_score') gt_a = dataset[i].get('gpt3_adherence') safe_append(gt_relevance, pred_relevance, gt_r, predicted['Context Relevance']) safe_append(gt_utilization, pred_utilization, gt_u, predicted['Context Utilization']) safe_append(gt_completeness, pred_completeness, gt_c, predicted['Completeness']) if gt_a is not None and predicted['Adherence'] is not None: safe_append(gt_adherence, pred_adherence, int(gt_a), int(predicted['Adherence'])) def compute_rmse(gt, pred): return round(np.sqrt(np.mean((np.array(gt) - np.array(pred)) ** 2)), 4) result = { "Context Relevance": compute_rmse(gt_relevance, pred_relevance), "Context Utilization": compute_rmse(gt_utilization, pred_utilization), "Completeness": compute_rmse(gt_completeness, pred_completeness), } if len(set(gt_adherence)) == 2: result["Adherence"] = compute_rmse(gt_adherence, pred_adherence) result["AUC-ROC (Adherence)"] = round(roc_auc_score(gt_adherence, pred_adherence), 4) else: result["Adherence"] = compute_rmse(gt_adherence, pred_adherence) result["AUC-ROC (Adherence)"] = "N/A - one class only" return result # Updated wrapper def evaluate_rag_gradio(domain, q_indices_str): # Capture logs log_stream = io.StringIO() sys.stdout = log_stream try: # Parse comma-separated indices q_indices = [int(x.strip()) for x in q_indices_str.split(",") if x.strip().isdigit()] results = evaluate_rag_pipeline(domain, q_indices) logs = log_stream.getvalue() return results, logs except Exception as e: traceback.print_exc() return {"error": str(e)}, log_stream.getvalue() finally: sys.stdout = sys.__stdout__ # Restore stdout # Gradio interface iface = gr.Interface( fn=evaluate_rag_gradio, inputs=[ gr.Dropdown(choices=["Legal", "Medical", "GK", "CS", "Finance"], label="Domain"), gr.Textbox(label="Comma-separated Query Indices (e.g. 89,121,245)", lines=1), ], outputs=[ gr.JSON(label="Evaluation Metrics (RMSE & AUC-ROC)"), gr.Textbox(label="Execution Log", lines=10, interactive=True), ], title="RAG Evaluation Dashboard", description="Evaluate your RAG pipeline across selected queries using GPT-based generation and judgment." ) # Launch app iface.launch(server_name="0.0.0.0", server_port=7860, debug=True)