Upload 3 files

README.md

@@ -1,11 +1,11 @@
 ---
-title: Federal Caselaw Index (with metadata-based reranker)
+title: NJ Caselaw Index Prototype
 emoji: 📈
 colorFrom: pink
 colorTo: green
 sdk: gradio
-sdk_version: 5.23.1
-app_file: app.py
+sdk_version: 5.0.2
+app_file: app_federal.py
 pinned: false
-short_description: Federal caselaw with metadata-based reranker
+short_description: NJ caselaw version 0, work in progress
 ---

app_federal.py

@@ -0,0 +1,291 @@
+import spaces
+import bm25s
+import gradio as gr
+import json
+import Stemmer
+import time
+import torch
+import os
+from transformers import AutoTokenizer, AutoModel, pipeline , AutoModelForSequenceClassification, AutoModelForCausalLM
+from sentence_transformers import SentenceTransformer
+import faiss
+import numpy as np
+import pandas as pd
+import torch.nn.functional as F
+from datasets import concatenate_datasets, load_dataset, load_from_disk
+from huggingface_hub import hf_hub_download
+from contextual import ContextualAI
+from openai import AzureOpenAI
+from datetime import datetime
+import sys
+from datetime import datetime
+from pathlib import Path
+from uuid import uuid4
+from huggingface_hub import CommitScheduler
+
+
+JSON_DATASET_DIR = Path("json_dataset")
+JSON_DATASET_DIR.mkdir(parents=True, exist_ok=True)
+JSON_DATASET_PATH = JSON_DATASET_DIR / f"train-{uuid4()}.json"
+
+scheduler = CommitScheduler(
+    repo_id="ai-law-society-lab/NJ-caselaw-queries",
+    repo_type="dataset",
+    folder_path=JSON_DATASET_DIR,
+    path_in_repo="data", token=os.getenv('hf_token')
+)
+
+
+def format_metadata_as_str(metadata):
+    try:
+        out =  metadata["case_name"] + ", " + metadata["court_short_name"] + ", " + metadata["date_filed"] + ", precedential status " + metadata["precedential_status"]
+    except:
+        out = ""
+    return out
+
+def show_user_query(user_message, history):
+    '''
+    Displays user query in the chatbot and removes from textbox.
+    :param user_message: user query inputted.
+    :param history: 2D array representing chatbot-user conversation.
+    :return:
+    '''
+    return "", history + [[user_message, None]]
+
+
+def run_extractive_qa(query, contexts):
+    extracted_passages = extractive_qa([{"question": query, "context": context} for context in contexts])        
+    return extracted_passages
+
+
+@spaces.GPU(duration=15)
+def respond_user_query(history):
+    '''
+    Overwrite the value of current pairing's history with generated text
+    and displays response character-by-character with some lag.
+    :param history: 2D array of chatbot history filled with user-bot interactions
+    :return: history updated with bot's latest message.
+    '''
+    start_time_global = time.time()
+
+    query = history[0][0]
+    start_time_global = time.time()
+
+    responses = run_retrieval(query)
+    print("--- run retrieval: %s seconds ---" % (time.time() - start_time_global))
+    #print (responses)
+
+    contexts = [individual_response["text"] for individual_response in responses][:NUM_RESULTS]
+    extracted_passages = run_extractive_qa(query, contexts)
+
+    for individual_response, extracted_passage in zip(responses, extracted_passages):
+        start, end = extracted_passage["start"], extracted_passage["end"]
+        # highlight text
+        text = individual_response["text"]
+        text = text[:start] + " **" + text[start:end] + "** " + text[end:]
+
+        # display queries in interface
+        formatted_response = "##### "
+        if individual_response["meta_data"]:
+            formatted_response += individual_response["meta_data"]
+        else:
+            formatted_response += individual_response["opinion_idx"]
+        formatted_response += "\n" + text + "\n\n"
+        history = history + [[None, formatted_response]]
+    print("--- Extractive QA: %s seconds ---" % (time.time() - start_time_global))
+
+    return [history, responses]
+
+def switch_to_reviewing_framework():
+    '''
+    Replaces textbox for entering user query with annotator review select.
+    :return: updated visibility for textbox and radio button props.
+    '''
+    return gr.Textbox(visible=False), gr.Dataset(visible=False), gr.Textbox(visible=True, interactive=True), gr.Button(visible=True)
+
+def reset_interface():
+    '''
+    Resets chatbot interface to original position where chatbot history,
+     reviewing is invisbile is empty and user input textbox is visible.
+    :return: textbox visibility, review radio button invisibility,
+    next_button invisibility, empty chatbot
+    '''
+
+    # remove tmp highlighted word documents
+    #for fn in os.listdir("tmp-docs"):
+    #    os.remove(os.path.join("tmp-docs", fn))
+    return gr.Textbox(visible=True), gr.Button(visible=False), gr.Textbox(visible=False, value=""), None, gr.JSON(visible=False, value=[]), gr.Dataset(visible=True)
+
+###################################################
+def mark_like(response_json, like_data: gr.LikeData):
+    index_of_msg_reviewed = like_data.index[0] - 1  # 0-indexing
+    # add liked information to res
+    response_json[index_of_msg_reviewed]["is_msg_liked"] = like_data.liked
+    return response_json
+
+"""
+def save_json(name: str, greetings: str) -> None:
+
+"""
+def register_review(history, additional_feedback, response_json):
+    '''
+    Writes user review to output file.
+    :param history: 2D array representing bot-user conversation so far.
+    :return: None, writes to output file.
+    ''' 
+
+    res = { "user_query": history[0][0], 
+        "responses": response_json,
+        "timestamp": datetime.now().strftime('%Y-%m-%d %H:%M:%S'), 
+        "additional_feedback": additional_feedback
+    }
+
+    with scheduler.lock:
+        with JSON_DATASET_PATH.open("a") as f:
+            json.dump(res, f)
+            f.write("\n")
+
+
+# load search functionality here
+
+
+def load_bm25():
+    stemmer = Stemmer.Stemmer("english")
+    retriever = bm25s.BM25.load("NJ_index_LLM_chunking", mmap=False)
+    return retriever, stemmer # titles
+
+def run_bm25(query):
+    query_tokens = bm25s.tokenize(query, stemmer=stemmer)
+    results, scores = retriever.retrieve(query_tokens, k=5)
+    return results[0]
+
+def load_faiss_index(embeddings):
+    nb, d = embeddings.shape # database size, dimension
+    faiss_index = faiss.IndexFlatL2(d)   # build the index
+    faiss_index.add(embeddings) # add vectors to the index
+    return faiss_index
+
+#@spaces.GPU(duration=10)
+def run_dense_retrieval(query):
+    if "NV" in model_name:
+        query_prefix = "Instruct: Given a question, retrieve passages that answer the question\nQuery: "
+        max_length = 32768
+        print (query)
+        with torch.no_grad():
+            query_embeddings = model.encode([query], instruction=query_prefix, max_length=max_length)
+            query_embeddings = F.normalize(query_embeddings, p=2, dim=1)
+        query_embeddings = query_embeddings.cpu().numpy()
+    return query_embeddings
+
+
+def load_NJ_caselaw():
+    if os.path.exists("/scratch/gpfs/ds8100/datasets/NJ_opinions_modernbert_splitter.jsonl"):
+        df = pd.read_json("/scratch/gpfs/ds8100/datasets/NJ_opinions_modernbert_splitter.jsonl", lines=True)
+    else:
+        df = pd.read_json("NJ_opinions_modernbert_splitter.jsonl", lines=True)
+    titles, chunks = [],[]
+
+    for i, row in df.iterrows():
+        texts = [i for i in row["texts"] if len(i.split()) > 25 and len(i.split()) < 750]
+        texts = [" ".join(i.strip().split()) for i in texts]
+        chunks.extend(texts)
+        titles.extend([row["id"]] * len(texts))
+    ids = list(range(len(titles)))
+    assert len(ids) == len(titles) == len(chunks)
+    return ids, titles, chunks
+
+
+def run_retrieval(query):
+    query = " ".join(query.split())
+    print ("query", query)
+
+    query_embeddings = run_dense_retrieval(query)
+    D, I = faiss_index.search(query_embeddings, 45)
+    scores_embeddings = D[0]
+    indices_embeddings = I[0]
+
+    #results = [{"index":i, "NV_score":j, "text": chunks[i]} for i,j in zip(indices_embeddings, scores_embeddings)]
+    results = [{"id":i, "NV_score":j, "text":ds_paragraphs[i]["paragraph"]} for i,j in zip(indices_embeddings, scores_embeddings)]
+
+    out_dict = []
+    covered = set()
+    for item in results:
+        index = item["index"]
+        item["query"] = query
+        item["opinion_idx"] = str(ds_paragraphs[i]["idx"])
+
+        # only recover one paragraph / opinion
+        if item["opinion_idx"] in covered:
+            continue
+        covered.add(item["opinion_idx"])
+
+        if item["opinion_idx"] in metadata:
+            item["meta_data"] = format_metadata_as_str(metadata[item["opinion_idx"]])
+        else:
+            item["meta_data"] = ""        
+        out_dict.append(item)
+    return out_dict
+
+
+NUM_RESULTS = 5
+model_name = 'nvidia/NV-Embed-v2' 
+
+#device = torch.device("cuda")
+device = torch.device("mps")
+
+extractive_qa = pipeline("question-answering", model="ai-law-society-lab/extractive-qa-model", tokenizer="FacebookAI/roberta-large", device_map="auto", token=os.getenv('hf_token'))
+
+ds_paragraphs = load_dataset("ai-law-society-lab/federal-caselaw-paragraphs", token=os.getenv('hf_token'))["train"]
+
+ds = load_dataset("ai-law-society-lab/federal-caselaw-embeddings-PCA", token=os.getenv('hf_token'))["train"]
+ds = ds.with_format("np")
+faiss_index = load_faiss_index(ds["embeddings"])
+
+
+with open("NJ_caselaw_metadata.json") as f:
+    metadata = json.load(f)
+ 
+
+def load_embeddings_model(model_name = "intfloat/e5-large-v2"):
+    if "NV" in model_name:
+        #model = AutoModel.from_pretrained('nvidia/NV-Embed-v2', trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="auto")
+        model = AutoModel.from_pretrained('nvidia/NV-Embed-v2', trust_remote_code=True, torch_dtype=torch.float16, device_map="auto")        
+        model.eval()
+        return model
+
+if "NV" in model_name:
+    model = load_embeddings_model(model_name=model_name)
+
+
+examples = ["Can officers always order a passenger out of a car?"]
+
+css = """
+.svelte-i3tvor {visibility: hidden}
+.row.svelte-hrj4a0.unequal-height {
+    align-items: stretch !important
+}
+"""
+
+with gr.Blocks(css=css, theme = gr.themes.Monochrome(primary_hue="pink",)) as demo:
+    chatbot = gr.Chatbot(height="45vw", autoscroll=False)
+    query_textbox = gr.Textbox()
+    #rerank_instruction = gr.Textbox(label="Rerank Instruction Prompt", value="If not otherwise specified in the query, prioritize Supreme Court opinions or opinions from higher courts. More recent, highly cited and published documents should also be weighted higher, unless otherwise specified in the query.")
+    examples = gr.Examples(examples, query_textbox)
+    response_json = gr.JSON(visible=False, value=[])
+    print (response_json)
+    chatbot.like(mark_like, response_json, response_json) 
+
+    feedback_textbox = gr.Textbox(label="Additional feedback?", visible=False)
+    next_button = gr.Button(value="Submit Feedback", visible=False)
+
+    query_textbox.submit(show_user_query, [query_textbox, chatbot], [query_textbox, chatbot], queue=False).then(
+        respond_user_query, chatbot, [chatbot, response_json]).then(
+        switch_to_reviewing_framework, None, [query_textbox, examples.dataset, feedback_textbox, next_button]
+    ) 
+
+    # Handle page reset and review save in database
+    next_button.click(register_review, [chatbot, feedback_textbox, response_json], None).then(
+        reset_interface, None, [query_textbox, next_button, feedback_textbox, chatbot, response_json, examples.dataset])
+
+# Launch application
+demo.launch() 

requirements.txt

@@ -2,7 +2,6 @@ spaces
 --extra-index-url https://download.pytorch.org/whl/cu113
 contextual-client
 openai
-gradio==5.23.1
 accelerate
 bm25s
 PyStemmer
@@ -13,4 +12,6 @@ sentence-transformers==2.7.0
 transformers==4.37.2
 einops
 numpy<2
-datasets
+datasets
+gradio
+pydantic==2.10.6