import os import gradio as gr import chromadb from sentence_transformers import SentenceTransformer import spaces client = chromadb.PersistentClient(path="./chroma") #collection_de = client.get_collection(name="phil_de") collection_en = client.get_collection(name="phil_en") #authors_list_de = ["Ludwig Wittgenstein", "Sigmund Freud", "Marcus Aurelius", "Friedrich Nietzsche", "Epiktet", "Ernst Jünger", "Georg Christoph Lichtenberg", "Balthasar Gracian", "Hannah Arendt", "Erich Fromm", "Albert Camus"] authors_list_en = ["Friedrich Nietzsche", "Joscha Bach"] @spaces.GPU def get_embeddings(queries, task): model = SentenceTransformer("Linq-AI-Research/Linq-Embed-Mistral", use_auth_token=os.getenv("HF_TOKEN")) prompts = [f"Instruct: {task}\nQuery: {query}" for query in queries] query_embeddings = model.encode(prompts) return query_embeddings def query_chroma(collection, embedding, authors): results = collection.query( query_embeddings=[embedding.tolist()], n_results=10, where={"author": {"$in": authors}} if authors else {}, include=["documents", "metadatas", "distances"] ) ids = results.get('ids', [[]])[0] metadatas = results.get('metadatas', [[]])[0] documents = results.get('documents', [[]])[0] distances = results.get('distances', [[]])[0] formatted_results = [] for id_, metadata, document_text, distance in zip(ids, metadatas, documents, distances): result_dict = { "id": id_, "author": metadata.get('author', 'Unknown author'), "book": metadata.get('book', 'Unknown book'), "section": metadata.get('section', 'Unknown section'), "title": metadata.get('title', 'Untitled'), "text": document_text, "distance": distance } formatted_results.append(result_dict) return formatted_results with gr.Blocks(css=".custom-markdown { border: 1px solid #ccc; padding: 10px; border-radius: 5px; }") as demo: gr.Markdown("Enter your query, filter authors (default is all), click **Search** to search. Delimit multiple queries with semicola; since there is a quota for each user (based on IP) it makes sense to query in batches. The search takes around 40 seconds, regardless of the number of queries, because the embedding model needs to be loaded to a GPU each time.") #database_inp = gr.Dropdown(label="Database", choices=["German", "English"], value="German") author_inp = gr.Dropdown(label="Authors", choices=authors_list_en, multiselect=True) inp = gr.Textbox(label="Query", placeholder="How can I live a healthy life?; How can I improve my ability to focus?; What is the meaning of life?; ...") btn = gr.Button("Search") results = gr.State() #def update_authors(database): # return gr.update(choices=authors_list_de if database == "German" else authors_list_en) #database_inp.change( # fn=lambda database: update_authors(database), # inputs=[database_inp], # outputs=[author_inp] #) def perform_query(queries, authors, database): task = "Given a question, retrieve passages that answer the question" queries = [query.strip() for query in queries.split(';')] embeddings = get_embeddings(queries, task) #collection = collection_de if database == "German" else collection_en collection = collection_en results_data = [] for query, embedding in zip(queries, embeddings): res = query_chroma(collection, embedding, authors) results_data.append((query, res)) return results_data btn.click( perform_query, inputs=[inp, author_inp], outputs=[results] ) @gr.render(inputs=[results]) def display_accordion(data): for query, res in data: with gr.Accordion(query, open=False) as acc: for result in res: with gr.Column(): author = result.get('author', 'Unknown author') book = result.get('book', 'Unknown book') text = result.get('text') markdown_contents = f"**{author}, {book}**\n\n{text}" gr.Markdown(value=markdown_contents, elem_classes="custom-markdown") demo.launch()