bug(python): hybrid search silently ignores .offset() #3765

Bug

The synchronous LanceHybridQueryBuilder (used by Table.search(query_type="hybrid") / .search(...).vector(...).text(...)) silently ignores .offset(). Calling .offset(N) on a hybrid query has no effect on the returned rows — the result is identical to .offset(0).

Root cause

In python/python/lancedb/query.py:

• LanceHybridQueryBuilder._create_query_builders() builds the internal self._vector_query / self._fts_query sub-builders and forwards limit, columns, where, etc., but never forwards self._offset to either sub-query.
• LanceHybridQueryBuilder._combine_hybrid_results() combines and reranks the two result sets and then does results.slice(length=limit) — always starting from position 0, regardless of any configured offset.

So self._offset (set via the inherited LanceQueryBuilder.offset()) is set but never read anywhere in the hybrid code path.

Note: the async path (AsyncHybridQuery, backed by the Rust HybridQuery::offset) does forward offset, but applies it independently to each sub-query before combining/reranking, which is a different (and arguably also incorrect) semantics than "skip the first N rows of the final reranked result." That's a separate, Rust-level concern and is out of scope for this issue, which is specifically about the sync Python path silently doing nothing.

Repro

import lancedb

db = lancedb.connect("memory://")
table = db.create_table(
    "t",
    [
        {"text": "a", "vector": [0.1, 0.1]},
        {"text": "b", "vector": [2, 2]},
        {"text": "cat", "vector": [-0.1, -0.1]},
        {"text": "dog", "vector": [0.5, -0.5]},
    ],
)
table.create_fts_index("text", with_position=False)

full = (
    table.search(query_type="hybrid")
    .vector([0.0, 0.4])
    .text("dog")
    .limit(4)
    .with_row_id()
    .to_arrow()
)

offset_result = (
    table.search(query_type="hybrid")
    .vector([0.0, 0.4])
    .text("dog")
    .offset(2)
    .limit(2)
    .with_row_id()
    .to_arrow()
)

# Expected: offset_result == full[2:4]
# Actual: offset_result == full[0:2] (offset is ignored)
print(full["_rowid"].to_pylist())
print(offset_result["_rowid"].to_pylist())

Environment

Reproduces on current main (confirmed by diffing the affected file against upstream — the bug is not fork-specific).

Fix

I've opened a PR that:

• Forwards self._offset into the per-sub-query fetch limit (limit + offset), so each sub-query fetches enough rows to cover the skipped prefix.
• Applies offset when slicing the final combined/reranked table (results.slice(offset=offset, length=limit)).
• Adds a regression test in python/python/tests/test_hybrid_query.py.