Spaces:
Build error
Build error
Download app.py from ml-usecases/media_assistant: direct link, hf CLI and curl.
- Browser
- Download file 6.35 kB
-
https://huggingface.co/spaces/ml-usecases/media_assistant/resolve/main/app.py
- Command line
-
hf download hf://spaces/ml-usecases/media_assistant/app.py
-
curl -L -o app.py https://huggingface.co/spaces/ml-usecases/media_assistant/resolve/main/app.py
6.35 kB
| """ | |
| Charity store assistant app | |
| """ | |
| import os | |
| from io import StringIO | |
| from datetime import datetime | |
| import pandas as pd | |
| from pandas import DataFrame | |
| import gradio as gr | |
| from pinecone import Pinecone | |
| import torch | |
| from transformers import AutoTokenizer, AutoModel | |
| from dotenv import load_dotenv | |
| import anthropic | |
| import boto3 | |
| TRANSFORMER_MODEL = "sentence-transformers/all-mpnet-base-v2" | |
| LLM = "claude-3-haiku-20240307" | |
| replace_dict = {"/'": "", '"': "", "'": "", "\n\n": " "} | |
| SIMILAR_MEDIA_LIST_LENGTH = 5 | |
| RESPONSE_WORD_LIMIT = 50 | |
| SYSTEM = f"You are helping to recommend music, movies or books based on a description. \ | |
| Your answer must be limited to {RESPONSE_WORD_LIMIT} words." | |
| pc = Pinecone(api_key=os.environ["PINECONE_KEY"]) | |
| index = pc.Index("product-recommender") | |
| tokenizer = AutoTokenizer.from_pretrained(TRANSFORMER_MODEL) | |
| model = AutoModel.from_pretrained(TRANSFORMER_MODEL) | |
| client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"]) | |
| def text_to_embedding(text: str): | |
| """ | |
| Convert text to vector embedding | |
| """ | |
| inputs = tokenizer(text, padding=True, truncation=True, return_tensors="pt") | |
| with torch.no_grad(): | |
| outputs = model(**inputs) | |
| embeddings = outputs.last_hidden_state.mean(dim=1).numpy() | |
| return embeddings.flatten().tolist() | |
| def find_similar_media(query_text: str, top_k=SIMILAR_MEDIA_LIST_LENGTH): | |
| """ | |
| Find media that is similar to the query | |
| """ | |
| query_embedding = text_to_embedding(query_text) | |
| matches = index.query( | |
| vector=query_embedding, | |
| top_k=top_k, | |
| include_metadata=True, | |
| filter={"vector_type": "test"}, | |
| )["matches"] | |
| df = pd.DataFrame( | |
| {"barcode": [m["id"] for m in matches], "score": [m["score"] for m in matches]} | |
| ) | |
| for col in ["artist", "title", "media_type", "description", "vector_type"]: | |
| df[col] = [m["metadata"][col] for m in matches] | |
| return df | |
| def get_llm_response(input_text: str) -> str: | |
| """ | |
| Retrieve formatted response from the LLM | |
| """ | |
| response = client.messages.create( | |
| model=LLM, | |
| max_tokens=1000, | |
| temperature=0, | |
| system=SYSTEM, | |
| messages=[{"role": "user", "content": [{"type": "text", "text": input_text}]}], | |
| ) | |
| response = response.content[0].text | |
| for k, v in replace_dict.items(): | |
| response = response.replace(k, v) | |
| return response | |
| def motivate_recommendation(title: str, artist: str, description: str, query_text: str): | |
| """ | |
| Ask the LLM to motivate why the most similar query matches the search string | |
| """ | |
| no_artist = artist == "(none)" | |
| if no_artist: | |
| prompt = ( | |
| f"Based on the description: '{description}', explain why \"{title} " | |
| f"matches the query: '{query_text}'." | |
| ) | |
| else: | |
| prompt = ( | |
| f'Based on the description: \'{description}\', explain why "{title}" by "{artist}" ' | |
| f"matches the query: '{query_text}'." | |
| ) | |
| response = get_llm_response(prompt) | |
| if no_artist: | |
| recommendation = ( | |
| f'May I suggest "{title}" by "{artist}"? It matches what you were ' | |
| f"looking for because {response}" | |
| ) | |
| else: | |
| recommendation = ( | |
| f'May I suggest "{title}"? It matches what you were ' | |
| f"looking for because {response}" | |
| ) | |
| return recommendation | |
| def log_dialogue(dialogue: dict): | |
| """ | |
| Log a user query and LLM response dialogue pair to a database | |
| """ | |
| session = boto3.Session() | |
| s3 = session.resource("s3") | |
| bucket_name, s3_file_name = "charityshop-media-assistant", "chat_history.csv" | |
| chat_history = pd.read_csv( | |
| StringIO( | |
| s3.Object(bucket_name, s3_file_name).get()["Body"].read().decode("utf-8") | |
| ) | |
| ) | |
| timestamp_str = datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f") | |
| new_dialogue = pd.DataFrame( | |
| { | |
| "store_name": [dialogue["ip_address"], dialogue["ip_address"]], | |
| "location": [dialogue["url"], dialogue["url"]], | |
| "recommendation_barcode": ["(none)", dialogue["recommendation_barcode"]], | |
| "role": ["user", "assistant"], | |
| "chat_text": [dialogue["user_query"], dialogue["motivation"]], | |
| "session_uuid": [dialogue["session_uuid"], dialogue["session_uuid"]], | |
| "timestamp": [timestamp_str, timestamp_str], | |
| } | |
| ) | |
| chat_history = pd.concat([chat_history, new_dialogue], ignore_index=True) | |
| csv_buffer = StringIO() | |
| chat_history.to_csv(csv_buffer, index=False) | |
| s3.Object(bucket_name, s3_file_name).put(Body=csv_buffer.getvalue()) | |
| def combined_recommendation( | |
| user_query: str, request: gr.Request | |
| ) -> tuple[str, DataFrame]: | |
| """ | |
| Wrapper function to receive user query and return | |
| the motivation for the best match and a table with the top 5 matches | |
| """ | |
| session_hash = request.session_hash if request else "(none)" | |
| ip_address = request.client.host if request else "(none)" | |
| url = request.url if request else "(none)" | |
| top_matches = find_similar_media(user_query) | |
| most_similar = top_matches.loc[0] | |
| motivation = motivate_recommendation( | |
| most_similar.title, most_similar.artist, most_similar.description, user_query | |
| ) | |
| dialogue = { | |
| "user_query": user_query, | |
| "motivation": motivation, | |
| "recommendation_barcode": top_matches.loc[0, "barcode"], | |
| "session_uuid": session_hash, | |
| "ip_address": ip_address, | |
| "url": url, | |
| } | |
| log_dialogue(dialogue) | |
| return motivation, top_matches[["media_type", "artist", "title"]] | |
| if __name__ == "__main__": | |
| load_dotenv() | |
| with gr.Blocks() as multi_output_app: | |
| input_textbox = gr.Textbox( | |
| show_label=False, placeholder="Type your query here..." | |
| ) | |
| submit_button = gr.Button("Submit") | |
| with gr.Tabs(): | |
| with gr.TabItem("Best Recommendation"): | |
| motivation_output = gr.Textbox(show_label=False) | |
| with gr.TabItem(f"Top {str(SIMILAR_MEDIA_LIST_LENGTH)} Matches"): | |
| table_output = gr.DataFrame(headers=["Result"]) | |
| submit_button.click( | |
| combined_recommendation, | |
| inputs=input_textbox, | |
| outputs=[motivation_output, table_output], | |
| ) | |
| multi_output_app.launch() | |