khoj/search_type/asymmetric.py

#!/usr/bin/env python

import json
from sentence_transformers import SentenceTransformer, CrossEncoder, util
import time
import gzip
import os
import sys
import re
import torch
import argparse
import pathlib
from utils.helpers import get_absolute_path

def initialize_model():
    "Initialize model for assymetric semantic search. That is, where query smaller than results"
    bi_encoder = SentenceTransformer('sentence-transformers/msmarco-MiniLM-L-6-v3')  # The bi-encoder encodes all entries to use for semantic search
    top_k = 100                                                                      # Number of entries we want to retrieve with the bi-encoder
    cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')             # The cross-encoder re-ranks the results to improve quality
    return bi_encoder, cross_encoder, top_k


def extract_entries(notesfile, verbose=0):
    "Load entries from compressed jsonl"
    entries = []
    with gzip.open(get_absolute_path(notesfile), 'rt', encoding='utf8') as jsonl:
        for line in jsonl:
            note = json.loads(line.strip())

            # Ignore title notes i.e notes with just headings and empty body
            if not "Body" in note or note["Body"].strip() == "":
                continue

            note_string = f'{note["Title"]}\t{note["Tags"] if "Tags" in note else ""}\n{note["Body"] if "Body" in note else ""}'
            entries.extend([note_string])

    if verbose > 0:
        print(f"Loaded {len(entries)} entries from {notesfile}")

    return entries


def compute_embeddings(entries, bi_encoder, embeddings_file, regenerate=False, verbose=0):
    "Compute (and Save) Embeddings or Load Pre-Computed Embeddings"
    # Load pre-computed embeddings from file if exists
    if embeddings_file.exists() and not regenerate:
        corpus_embeddings = torch.load(get_absolute_path(embeddings_file))
        if verbose > 0:
            print(f"Loaded embeddings from {embeddings_file}")

    else:  # Else compute the corpus_embeddings from scratch, which can take a while
        corpus_embeddings = bi_encoder.encode(entries, convert_to_tensor=True, show_progress_bar=True)
        torch.save(corpus_embeddings, get_absolute_path(embeddings_file))
        if verbose > 0:
            print(f"Computed embeddings and save them to {embeddings_file}")

    return corpus_embeddings


def query_notes(raw_query, corpus_embeddings, entries, bi_encoder, cross_encoder, top_k=100):
    "Search all notes for entries that answer the query"
    # Separate natural query from explicit required, blocked words filters
    query = " ".join([word for word in raw_query.split() if not word.startswith("+") and not word.startswith("-")])
    required_words = set([word[1:].lower() for word in raw_query.split() if word.startswith("+")])
    blocked_words = set([word[1:].lower() for word in raw_query.split() if word.startswith("-")])

    # Encode the query using the bi-encoder
    question_embedding = bi_encoder.encode(query, convert_to_tensor=True)

    # Find relevant entries for the query
    hits = util.semantic_search(question_embedding, corpus_embeddings, top_k=top_k)
    hits = hits[0]  # Get the hits for the first query

    # Filter results using explicit filters
    hits = explicit_filter(hits, entries, required_words, blocked_words)
    if hits is None or len(hits) == 0:
        return hits

    # Score all retrieved entries using the cross-encoder
    cross_inp = [[query, entries[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)

    # Store cross-encoder scores in results dictionary for ranking
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]

    # Order results by cross encoder score followed by biencoder score
    hits.sort(key=lambda x: x['score'], reverse=True) # sort by biencoder score
    hits.sort(key=lambda x: x['cross-score'], reverse=True) # sort by cross encoder score

    return hits


def explicit_filter(hits, entries, required_words, blocked_words):
    hits_by_word_set = [(set(word.lower()
                             for word
                             in re.split(
                                 ',|\.| |\]|\[\(|\)|\{|\}',
                                 entries[hit['corpus_id']])
                             if word != ""),
                         hit)
                        for hit in hits]

    if len(required_words) == 0 and len(blocked_words) == 0:
        return hits
    if len(required_words) > 0:
        return [hit for (words_in_entry, hit) in hits_by_word_set
                if required_words.intersection(words_in_entry) and not blocked_words.intersection(words_in_entry)]
    if len(blocked_words) > 0:
        return [hit for (words_in_entry, hit) in hits_by_word_set
                if not blocked_words.intersection(words_in_entry)]
    return hits


def render_results(hits, entries, count=5, display_biencoder_results=False):
    "Render the Results returned by Search for the Query"
    if display_biencoder_results:
        # Output of top hits from bi-encoder
        print("\n-------------------------\n")
        print(f"Top-{count} Bi-Encoder Retrieval hits")
        hits = sorted(hits, key=lambda x: x['score'], reverse=True)
        for hit in hits[0:count]:
            print(f"Score: {hit['score']:.3f}\n------------\n{entries[hit['corpus_id']]}")

    # Output of top hits from re-ranker
    print("\n-------------------------\n")
    print(f"Top-{count} Cross-Encoder Re-ranker hits")
    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
    for hit in hits[0:count]:
        print(f"CrossScore: {hit['cross-score']:.3f}\n-----------------\n{entries[hit['corpus_id']]}")


def collate_results(hits, entries, count=5):
    return [
        {
            "Entry": entries[hit['corpus_id']],
            "Score": f"{hit['cross-score']:.3f}"
        }
        for hit
        in hits[0:count]]


if __name__ == '__main__':
    # Setup Argument Parser
    parser = argparse.ArgumentParser(description="Map Org-Mode notes into (compressed) JSONL format")
    parser.add_argument('--compressed-jsonl', '-j', required=True, type=pathlib.Path, help="Input file for compressed JSONL formatted notes to compute embeddings from")
    parser.add_argument('--embeddings', '-e', required=True, type=pathlib.Path, help="File to save/load model embeddings to/from")
    parser.add_argument('--results-count', '-n', default=5, type=int, help="Number of results to render. Default: 5")
    parser.add_argument('--interactive', action='store_true', default=False, help="Interactive mode allows user to run queries on the model. Default: true")
    parser.add_argument('--verbose', action='count', default=0, help="Show verbose conversion logs. Default: 0")
    args = parser.parse_args()

    # Initialize Model
    bi_encoder, cross_encoder, top_k = initialize_model()

    # Extract Entries
    entries = extract_entries(args.compressed_jsonl, args.verbose)

    # Compute or Load Embeddings
    corpus_embeddings = compute_embeddings(entries, bi_encoder, args.embeddings, args.verbose)

    # Run User Queries on Entries in Interactive Mode
    while args.interactive:
        # get query from user
        user_query = input("Enter your query: ")
        if user_query == "exit":
            exit(0)

        # query notes
        hits = query_notes(user_query, corpus_embeddings, entries, bi_encoder, cross_encoder, top_k)

        # render results
        render_results(hits, entries, count=args.results_count)