barrel_ngram_verify (barrel_ngram v0.9.0)

Windowed literal and regex verification.

A planner candidate match start is necessary, never sufficient. windowed/4 confirms a literal candidate by reading just byte_size(Literal) bytes at its start; windowed_regex/7 reads the window barrel_ngram_planner:regex_plan/2 computed around a chosen anchor and re-runs the pattern over that slice. Both are case-sensitive only -- phase-2 sampling is itself case-sensitive, so a caseless query never reaches this module.

Both reduce to the same non-overlapping matches a plain left-to-right scan (binary:matches/2, re:runglobal) reports: two independently distance-checked candidates can be real but overlapping matches ("aaa" at both offset 0 and 1 of "aaaa"), and without the reduction this module would report both, diverging from every other lane. See non_overlapping/1.

Summary

Functions

non_overlapping(Spans)

Spans reduced to the same leftmost-greedy non-overlapping set a left-to-right scan produces: sorted by start, keep the earliest available span and skip anything starting before it ends.

windowed(Source, Key, Literal, Starts)

The subset of Starts where Source's bytes for Key, read at that start for byte_size(Literal) bytes, equal Literal exactly. A failed read (deleted document, or a race) drops that candidate silently.

windowed_regex(Source, Key, RE, PrefixMax, AnchorLen, SuffixMax, AnchorStarts)

The real {Start, Length} matches of RE, found by reading only the window around each AnchorStart in AnchorStarts: [AnchorStart - PrefixMax, AnchorStart + AnchorLen + SuffixMax), clamped to [0, doc_size). Spans are translated back to absolute offsets and deduplicated (more than one anchor can rediscover the same match). A failed doc_size/pread (deleted document, or a race) drops that document or window silently.

Functions

non_overlapping(Spans)

-spec non_overlapping([{non_neg_integer(), non_neg_integer()}]) ->
                         [{non_neg_integer(), non_neg_integer()}].

Spans reduced to the same leftmost-greedy non-overlapping set a left-to-right scan produces: sorted by start, keep the earliest available span and skip anything starting before it ends.

windowed(Source, Key, Literal, Starts)

-spec windowed({module(), term()}, binary(), binary(), [non_neg_integer()]) -> [non_neg_integer()].

windowed_regex(Source, Key, RE, PrefixMax, AnchorLen, SuffixMax, AnchorStarts)

-spec windowed_regex({module(), term()},
                     binary(),
                     re:mp(),
                     non_neg_integer(),
                     non_neg_integer(),
                     non_neg_integer(),
                     [non_neg_integer()]) ->
                        [{non_neg_integer(), non_neg_integer()}].