Duplicate search results when `k` is a high value about colbert

[okhat]

This is very unusual, it should not return repeated passages unless there’s some recent change causing that.

do you have a guess @santhnm2 @detaos @jessiejuachon

[santhnm2]

Agree with @okhat, this is very strange - I'll look into it.

[s-jse]

I'm running into another issue with large k values.
Running search with any large k, e.g. 100, returns exactly 64 passages.

[thesillypeanut]

@s-jse I've experienced that issue as well. The number of passages returned seems to get stuck at a specific number (a power of 2) no matter what value is passed. This was noticed to have happened whenever we vary the k value in a for-loop and it freezes at a specific number beyond which it doesn't return results even if they exist. Code snippet used to test this:

def get_answers(query):

    for k in [1, 10, 100]:
        print(f"Question: {query} k: {k}")
        results = searcher.search(query, k=k)
        print(f"NUM RESULTS: {len(results[0])}")

        unique_pids = set()
        for passage_id, passage_rank, passage_score in zip(*results):
            print(f"passage_id={passage_id} passage_rank={passage_rank} passage_score={passage_score}")
            unique_pids.add(passage_id)

        print(f"UNIQUE PIDS: {len(unique_pids)}")
        sleep(2)


if __name__ == '__main__':

    searcher, output_indexes, output_content = load_model()
    narratives = load_dataset()

    for i in range(100):

        narrative_dict = narratives[i]
        get_answers(narrative_dict['question'])

[paul7Junior]

I used Colbert for quite a lot those past days and I fixed this one for myself.
So I wanted to share a solution. I submitted a PR.
It's my first time submitting to an open source project and welcome any feedbacks.

Cheers

[okhat]

@paul7Junior really appreciate this! We’ll take a look!

[ravi-kumar-1010]

Hello! I tried to search using a large k value. I noticed that ColBERT returns some unique results, but when there aren't enough results, it soon starts returning duplicate results (with the same passage ID) until it reaches the k value specified.

For example, the code below:

query = "some question"   # or supply your own query

print(f"Question: {query}")
results = searcher.search(query, k=1000)
print(f"NUM RESULTS {len(results[0])}")

unique_pids = set()
repeated_pids = []

for passage_id, passage_rank, passage_score in zip(*results):
    text = searcher.collection[passage_id]

    if passage_id in unique_pids:
        repeated_pids.append(passage_id)
        continue

    unique_pids.add(passage_id)

print(len(repeated_pids), repeated_pids)
print(len(unique_pids), unique_pids)

prints the following:

Question: some question
NUM RESULTS 1000
872 [25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, ...]  # seems to duplicate just one passage over and over again
128 {1, 1027, 1034, 2572, 1041, 1043, ...}

Shouldn't ColBERT only be returning unique results? Is this is a known bug?

User code below, I was facing the same Issue, but using this custom_config object I was able to get 100 results or more.
custom_config = ColBERTConfig(ncells=1000, ndocs=1000,reranker=True)
with Run().context(RunConfig(experiment='notebook')):
searcher = Searcher(index=index_name,config=custom_config)

[paul7Junior]

Hey Ravi Kumar,

Thanks for your comment,

Can you just post here the version of colbert-ai that you're using.

Also, when you say you was able to get 100+ results, are you talking about 100+ results of the same passage id?