Spaces:
Running
Running
| import time | |
| import gradio as gr | |
| from datasets import load_dataset | |
| import pandas as pd | |
| from sentence_transformers import SentenceTransformer | |
| from sentence_transformers.quantization import quantize_embeddings | |
| import faiss | |
| from usearch.index import Index | |
| # Load titles and texts | |
| title_text_dataset = load_dataset("mixedbread-ai/wikipedia-data-en-2023-11", split="train", num_proc=4).select_columns(["title", "text"]) | |
| # Load the int8 and binary indices. Int8 is loaded as a view to save memory, as we never actually perform search with it. | |
| int8_view = Index.restore("wikipedia_int8_usearch_50m.index", view=True) | |
| binary_index: faiss.IndexBinaryFlat = faiss.read_index_binary("wikipedia_ubinary_faiss_50m.index") | |
| # Load the SentenceTransformer model for embedding the queries | |
| model = SentenceTransformer( | |
| "mixedbread-ai/mxbai-embed-large-v1", | |
| prompts={ | |
| "retrieval": "Represent this sentence for searching relevant passages: ", | |
| }, | |
| default_prompt_name="retrieval", | |
| ) | |
| def search(query, top_k: int = 10, rescore_multiplier: int = 4): | |
| # 1. Embed the query as float32 | |
| start_time = time.time() | |
| query_embedding = model.encode(query) | |
| embed_time = time.time() - start_time | |
| # 2. Quantize the query to ubinary | |
| start_time = time.time() | |
| query_embedding_ubinary = quantize_embeddings(query_embedding.reshape(1, -1), "ubinary") | |
| quantize_time = time.time() - start_time | |
| # 3. Search the binary index | |
| start_time = time.time() | |
| _scores, binary_ids = binary_index.search(query_embedding_ubinary, top_k * rescore_multiplier) | |
| binary_ids = binary_ids[0] | |
| search_time = time.time() - start_time | |
| # 4. Load the corresponding int8 embeddings | |
| start_time = time.time() | |
| int8_embeddings = int8_view[binary_ids].astype(int) | |
| load_time = time.time() - start_time | |
| # 5. Rescore the top_k * rescore_multiplier using the float32 query embedding and the int8 document embeddings | |
| start_time = time.time() | |
| scores = query_embedding @ int8_embeddings.T | |
| rescore_time = time.time() - start_time | |
| # 6. Sort the scores and return the top_k | |
| start_time = time.time() | |
| indices = scores.argsort()[:top_k] | |
| top_k_indices = binary_ids[indices] | |
| top_k_scores = scores[indices] | |
| top_k_titles, top_k_texts = zip(*[(title_text_dataset[idx]["title"], title_text_dataset[idx]["text"]) for idx in top_k_indices.tolist()]) | |
| df = pd.DataFrame({"Score": [round(value, 2) for value in top_k_scores], "Title": top_k_titles, "Text": top_k_texts}) | |
| sort_time = time.time() - start_time | |
| return df, { | |
| "Embed Time": f"{embed_time:.4f} s", | |
| "Quantize Time": f"{quantize_time:.4f} s", | |
| "Search Time": f"{search_time:.4f} s", | |
| "Load Time": f"{load_time:.4f} s", | |
| "Rescore Time": f"{rescore_time:.4f} s", | |
| "Sort Time": f"{sort_time:.4f} s", | |
| "Total Retrieval Time": f"{quantize_time + search_time + load_time + rescore_time + sort_time:.4f} s" | |
| } | |
| with gr.Blocks(title="Quantized Retrieval") as demo: | |
| gr.Markdown( | |
| """ | |
| ## Quantized Retrieval - Binary Search with Scalar (int8) Rescoring | |
| This demo showcases exact retrieval using [quantized embeddings](https://huggingface.co/blog/embedding-quantization). The corpus consists of 41 million texts from Wikipedia articles. | |
| <details><summary>Click to learn about the retrieval process</summary> | |
| Details: | |
| 1. The query is embedded using the [`mixedbread-ai/mxbai-embed-large-v1`](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) SentenceTransformer model. | |
| 2. The query is quantized to binary using the `quantize_embeddings` function from the SentenceTransformers library. | |
| 3. A binary index (41M binary embeddings; 5.2GB of memory/disk space) is searched using the quantized query for the top 40 documents. | |
| 4. The top 40 documents are loaded on the fly from an int8 index on disk (41M int8 embeddings; 0 bytes of memory, 47.5GB of disk space). | |
| 5. The top 40 documents are rescored using the float32 query and the int8 embeddings to get the top 10 documents. | |
| 6. The top 10 documents are sorted by score and displayed. | |
| This process is designed to be memory efficient and fast, with the binary index being small enough to fit in memory and the int8 index being loaded as a view to save memory. | |
| In total, this process requires keeping 1) the model in memory, 2) the binary index in memory, and 3) the int8 index on disk. With a dimensionality of 1024, | |
| we need `1024 / 8 * num_docs` bytes for the binary index and `1024 * num_docs` bytes for the int8 index. | |
| This is notably cheaper than doing the same process with float32 embeddings, which would require `4 * 1024 * num_docs` bytes of memory/disk space for the float32 index, i.e. 32x as much memory and 4x as much disk space. | |
| Additionally, the binary index is much faster (up to 32x) to search than the float32 index, while the rescoring is also extremely efficient. In conclusion, this process allows for fast, scalable, cheap, and memory-efficient retrieval. | |
| Feel free to check out the [code for this demo](https://huggingface.co/spaces/tomaarsen/quantized_retrieval/blob/main/app.py) to learn more about how to apply this in practice. | |
| </details> | |
| """) | |
| query = gr.Textbox(label="Query for Wikipedia articles", placeholder="Enter a query to search for relevant texts from Wikipedia.") | |
| search_button = gr.Button(value="Search") | |
| with gr.Row(): | |
| with gr.Column(scale=4): | |
| output = gr.Dataframe(headers=["Score", "Title", "Text"]) | |
| with gr.Column(scale=1): | |
| json = gr.JSON() | |
| query.submit(search, inputs=[query], outputs=[output, json]) | |
| search_button.click(search, inputs=[query], outputs=[output, json]) | |
| demo.queue() | |
| demo.launch(debug=True) | |