Skip to content
mdfu

Backend Evaluation

Question: stay indexless (sahilm/fuzzy + hand filters) or adopt an embedded full-text library for free-text + fuzzy + faceted custom metadata?

Why bleve was the prior recommendation

Section titled “Why bleve was the prior recommendation”

github.com/blevesearch/bleve/v2 (11k★, Apache-2.0, Couchbase-backed) is the only pure-Go option covering all three out of box:

  • Free text: text fields + analyzers, BM25/TF-IDF, highlighting, scorch on-disk or NewMemOnly in-memory, no cgo.
  • Fuzzy: NewFuzzyQuery (Levenshtein) + prefix/regexp/wildcard/match_phrase/query-string.
  • Facets/custom metadata: dynamic mapping indexes Raw.author/views/… automatically; terms/numeric-range/date-range facets; BooleanQuery(must: facets, should: fuzzy) combines filter + search.

Why not bluge?

Section titled “Why not bluge?”

blugelabs/bluge (+ zincsearch/blugehq/fy0 forks) is lighter and faster at indexing with a cleaner API and richer aggregations (Terms/Range/Min/Max/HLL/T-Digest). Reasons to hold off:

  • Fragmented ecosystem — unclear canonical fork, near-zero importers, thin docs.
  • Mainline query types lack fuzzy/regexp/wildcard (Term/Phrase/Match/Prefix/Range/Boolean only); FuzzyQuery exists only in forks, unproven.
  • Pick bluge only if write throughput + readable codebase outweigh fuzzy + facet maturity + community.

Why not SQLite FTS5?

Section titled “Why not SQLite FTS5?”

FTS5(porter+trigram) + normal columns gives the fastest indexing and best SQL facets (WHERE type=? AND tag=? GROUP BY). Costs:

  • No native fuzzy — needs spellfix1 two-step queries or trigram substring hacks.
  • Facets/ranking are hand-written SQL, not a search API.
  • Go drivers: mattn/go-sqlite3 needs cgo (breaks static binary/cross-compile); modernc.org/sqlite is pure-Go but larger/slower; FTS↔content sync (triggers/external-content) is extra work.
  • Pick FTS5 only if you already need SQLite or persistent SQL joins.

Decision for mdfu

Section titled “Decision for mdfu”
  • Stay indexless while the <50ms/keystroke budget holds (see Search & Ranking bottleneck + Roadmap optimization tasks).
  • Spike a persistent index only if benchmarks miss: bleve-mem first (facets UI), bluge for speed comparison, FTS5 for SQL — decide on binary size, cgo, cold-start, query latency.
  • External servers (Typesense/Meilisearch) are out of scope: violate the single-static-binary constraint.