wiki_funcs.py

from langchain.agents import tool 

from torch import tensor as torch_tensor
from datasets import load_dataset
from sentence_transformers import SentenceTransformer, CrossEncoder, util

"""# import models"""

bi_encoder = SentenceTransformer(
    'sentence-transformers/multi-qa-MiniLM-L6-cos-v1')
bi_encoder.max_seq_length = 256  # Truncate long passages to 256 tokens

# The bi-encoder will retrieve top_k documents. We use a cross-encoder, to re-rank the results list to improve the quality
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')

"""# import datasets"""
dataset = load_dataset("gfhayworth/wiki_mini", split='train')

mypassages = list(dataset.to_pandas()['psg'])

dataset_embed = load_dataset("gfhayworth/wiki_mini_embed", split='train')

dataset_embed_pd = dataset_embed.to_pandas()
mycorpus_embeddings = torch_tensor(dataset_embed_pd.values)


def search(query, top_k=20, top_n=1):
    question_embedding = bi_encoder.encode(query, convert_to_tensor=True)
    hits = util.semantic_search(
        question_embedding, mycorpus_embeddings, top_k=top_k)
    hits = hits[0]  # Get the hits for the first query

    ##### Re-Ranking #####
    cross_inp = [[query, mypassages[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)

    # Sort results by the cross-encoder scores
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]

    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
    predictions = hits[:top_n]
    return predictions
    # for hit in hits[0:3]:
    #     print("\t{:.3f}\t{}".format(hit['cross-score'], mypassages[hit['corpus_id']].replace("\n", " ")))


def get_text(qry):
    # predictions = greg_search(qry)
    predictions = search(qry)
    prediction_text = []
    for hit in predictions:
        prediction_text.append("{}".format(mypassages[hit['corpus_id']]))
    return prediction_text


@tool
def mysearch(query: str) -> str:
    """Query our own datasets.
    """
    rslt = get_text(query)
    return '\n'.join(rslt)


@tool
def mygreetings(greeting: str) -> str:
    """Let us do our greetings
    """

    return "how are you?"