[Schema streaming mode] bm25 score is always zero

xansrnitu commented 7 months ago

Describe the bug I am always getting bm25 score as zero even though the target text field contains my search key.

To Reproduce Steps to reproduce the behavior:

My Schema and ranking definition (python)-


Schema(
            name="summaries",
            mode="streaming", 
            document=VespaDoc(
                    fields=[
                        Field(name="id", type="string", indexing=["summary", "index"]),
                        Field(name="title", type="string", indexing=["summary", "index"]),
                        Field(name="page", type="int", indexing=["summary", "index"]),
                        Field(name="metadata", type="map<string,string>", indexing=["summary", "index"]),
                        Field(name="content", type="string", indexing=["summary", "index"]),
                        Field(name="embedding", type="tensor<bfloat16>(x[384])",
                            indexing=["input content", "embed e5", "attribute", "index"],
                            ann=HNSW(distance_metric="angular"),
                            is_document_field=False
                        )
                    ]
            ),
            fieldsets=[
                FieldSet(name = "default", fields = ["content"])
            ]
        )

     bm25 = RankProfile(
        name="bm25", 
        functions=[Function(
            name="my_bm25", expression="bm25(content)"
        )],
        first_phase=FirstPhaseRanking(
            expression="my_bm25"
        ),
        match_features=["my_bm25"],
    )   

    passage_schema.add_rank_profile(bm25)

Now, if I query with -

app.query(
                yql="select content from summaries where userQuery()",
                groupname="xyz",
                query="Jaganmohan",
                hits = 4,
                ranking="bm25"
            )

returns 'features': {'my_bm25': 0.0} despite the fact that "content" field contains "Jaganmohan". I have tried with some other queries as well, all giving 0 score.

Environment (please complete the following information):

OS:Linux Ubuntu
Infrastructure:docker - self-hosted

Vespa version 8.279.6

Additional context nativeRank(content) is returning positive score ( > 0)

bratseth commented 7 months ago

Thanks for reporting. The bm25 feature is correct during ranking (e.g if you output it as the first-phase score), but becomes 0 in sumary-features/match-features.

geirst commented 7 months ago

This is fixed in Vespa 8.302.40. System tests added in https://github.com/vespa-engine/system-test/pull/3322 and https://github.com/vespa-engine/system-test/pull/3327.

vespa-engine / vespa

[Schema streaming mode] bm25 score is always zero #30220