docs: state the path-table/declared-table trade-off, ship FTS5 and vector index templates

ARCHITECTURE.md

@@ -31,6 +31,55 @@ prior `[table.columns]` source-dispatch and the `format` / `each` config
 keys were ripped out; the per-format parser zoo (`Format::Json`, `Csv`,
 `Yaml`, `Toml`, `Frontmatter`, …) is gone.
 
+## Two table kinds, opposite trade-offs
+
+The two kinds above are not a general mechanism and a limited one. They are a
+choice the user makes per table, and each buys what the other gives up.
+
+A **path-table** costs nothing to reach for. `FROM './**/*.json'` needs no
+config file, no DDL and no hook; the table is minted in `temp` when the
+statement runs, so it lives and dies with the connection. Nothing is stored:
+the filesystem is walked during the query and `content` is read only if a
+column names it. What you get for that is freshness with no machinery — a file
+written a second ago is in the next statement's results, because there is no
+copy of the tree to go stale.
+
+A **declared table** costs a `[[table]]` entry and an `on-file` hook. In
+exchange it is an ordinary SQLite table: rows are inserted during the build,
+the watcher keeps them current as files change, `--persist` carries them
+across restarts, and its `ddl` batch can create anything SQLite can — a
+`CREATE INDEX`, an FTS5 vtab, a `vec0` vtab, and the triggers that keep them
+current.
+
+The asymmetry is structural, not an unfinished path-table. Indexes,
+persistence and the watcher all need rows that outlive the statement that
+asked for them, and a path-table deliberately has none — it stores nothing,
+which is exactly why it is always current and needs no setup. A table cannot
+both hold no state and index its state.
+
+So the trade reads:
+
+|  | Path-table | Declared table |
+| --- | --- | --- |
+| Setup | none | `[[table]]` entry + `on-file` hook |
+| Cost per query | walks the matched tree | reads rows from SQLite |
+| Freshness | always current | watcher-maintained |
+| Survives restart | no | with `--persist` |
+| Indexes (B-tree, FTS5, vector) | no | in the `ddl` batch |
+
+Reach for a path-table when the question is ad hoc, or the tree is small
+enough that walking it is not worth avoiding. Declare a table when the same
+tree is queried repeatedly, is large, or wants keyword or vector search. The
+two are ordinary SQLite tables once resolved, so a query is free to join one
+against the other.
+
+The user-facing versions of both sides are
+[Query files without a config](https://thekevinscott.github.io/dirsql/howto/query-without-config)
+and
+[Define tables for your files](https://thekevinscott.github.io/dirsql/howto/define-tables);
+the index templates are in
+[Add a search index to a table](https://thekevinscott.github.io/dirsql/howto/search-indexes).
+
 ## Read-only by design
 
 **dirsql never modifies the directory it indexes.** It opens files for reading
@@ -165,7 +214,7 @@ reinvent what SQLite expresses natively*. `validate_identifier` remains: `name`
 is spliced into `format!()`-built INSERT/DELETE SQL, so it is still the
 injection guard.
 
-**Named tables are real; path-tables are virtual.** A declared `[[table]]` (or programmatic `Table`) is a real SQLite table whose rows are inserted on build and maintained by the watcher — the `db` module above. A [path-table](../docs/reference/path-tables.md) (`SELECT * FROM './'`, epic path-as-table) is a `dirsql_path` **virtual table** (`vtab.rs` / `path_table.rs`, rusqlite's `vtab` feature): no rows are stored, no reconcile or watcher runs, and the filesystem is walked live at query time (`xFilter`/`xNext` enumerate matched files; `xColumn` supplies stat values and lazily reads `content`). SQLite stays the entire query engine either way; the vtab only enumerates rows and supplies column values. Path-tables are registered on demand — see *Query execution* below.
+**Named tables are real; path-tables are virtual.** A declared `[[table]]` (or programmatic `Table`) is a real SQLite table whose rows are inserted on build and maintained by the watcher — the `db` module above. A [path-table](https://thekevinscott.github.io/dirsql/reference/path-tables) (`SELECT * FROM './'`, epic path-as-table) is a `dirsql_path` **virtual table** (`vtab.rs` / `path_table.rs`, rusqlite's `vtab` feature): no rows are stored, no reconcile or watcher runs, and the filesystem is walked live at query time (`xFilter`/`xNext` enumerate matched files; `xColumn` supplies stat values and lazily reads `content`). SQLite stays the entire query engine either way; the vtab only enumerates rows and supplies column values. Path-tables are registered on demand — see *Query execution* below.
 
 ### `scanner` -- Directory traversal
 

docs/.vitepress/config.ts

@@ -56,6 +56,7 @@ export default defineConfig({
             { text: 'Parse your files into columns', link: '/howto/parse-files-into-columns' },
             { text: 'Query JSON file contents', link: '/howto/query-json' },
             { text: 'Search documents by meaning', link: '/howto/search-by-meaning' },
+            { text: 'Add a search index to a table', link: '/howto/search-indexes' },
             { text: "Skip files you don't want indexed", link: '/howto/skip-files' },
             { text: 'Load a SQLite extension', link: '/howto/load-extension' },
             { text: 'Keep the index across restarts', link: '/howto/persist' },

docs/howto/define-tables.md

@@ -107,7 +107,9 @@ maintain nothing. `name` still has to be the row table (`posts`); the virtual
 table lives beside it under its own name. The batch runs once, when the table
 is created, and editing any part of it rebuilds a
 [persistent cache](./persist.md) from scratch. Full rules:
-[Batch `ddl`](../reference/config.md#batch-ddl).
+[Batch `ddl`](../reference/config.md#batch-ddl); pasteable FTS5 and vector
+templates, with the trigger mistakes that fail silently, are in
+[Add a search index to a table](./search-indexes.md).
 
 ## Going further
 

docs/howto/query-without-config.md

@@ -55,7 +55,9 @@ question over a few hundred files, wasteful for a large tree you query
 repeatedly. When that day comes, [declare a table](./define-tables.md): it is
 indexed once, kept fresh by the watcher, and can persist across restarts. The
 zero-config query is the floor; a named table is the escalation path — nothing
-you write here has to be thrown away to get there.
+you write here has to be thrown away to get there. The two sides in full, and
+why indexes belong to only one of them, are in
+[how `dirsql` thinks](../explanation.md#two-table-kinds-opposite-trade-offs).
 
 ## Notes
 

docs/howto/search-indexes.md

@@ -0,0 +1,220 @@
+# Add a search index to a table
+
+Make a declared table answer search queries instead of scanning: a B-tree for
+lookups, an FTS5 index for keywords, a `vec0` index for meaning. All three are
+statements in the table's [`ddl` batch](../reference/config.md#batch-ddl), so
+SQLite does the work and `dirsql` maintains nothing.
+
+The templates below are meant to be pasted and edited. Both are shaped around
+one worked example — notes in `notes/*.md`, parsed into a `notes` table by an
+`extract.py` that prints `{"slug": …, "title": …, "body": …}` per file, exactly
+as in [Define tables for your files](./define-tables.md).
+
+## Why triggers, and why only two
+
+`dirsql` writes file rows with plain `INSERT` and `DELETE`; an update is a
+delete and an insert in one transaction, and there is no `UPDATE` path on user
+rows. So an insert trigger and a delete trigger cover every way a row can
+change — during the initial build, on each [watcher](./react-to-changes.md)
+event, and when a [persistent cache](./persist.md) reconciles a tree that
+moved on. There is no third trigger to write and no maintenance command to run.
+
+The triggers are also the part that goes wrong, which is why the templates
+spell them out rather than hiding them behind an abstraction.
+
+## Keyword search (FTS5)
+
+FTS5 ships inside SQLite — no extension, no install. Declare an
+external-content index beside the row table and let the two triggers feed it:
+
+```toml
+[[table]]
+name    = "notes"
+glob    = "notes/**/*.md"
+on-file = "python3 extract.py {path}"
+ddl     = '''
+CREATE TABLE notes (slug TEXT, title TEXT, body TEXT);
+
+CREATE VIRTUAL TABLE notes_fts USING fts5(
+  body,
+  content='notes',
+  content_rowid='rowid',
+  tokenize='porter unicode61'
+);
+CREATE TRIGGER notes_ai AFTER INSERT ON notes BEGIN
+  INSERT INTO notes_fts(rowid, body) VALUES (new.rowid, new.body);
+END;
+CREATE TRIGGER notes_ad AFTER DELETE ON notes BEGIN
+  INSERT INTO notes_fts(notes_fts, rowid, body)
+    VALUES ('delete', old.rowid, old.body);
+END;
+'''
+```
+
+`content='notes'` makes the index store only its own search structures and
+read column values back from `notes`, so the text is not duplicated.
+`tokenize='porter unicode61'` stems English on top of the default
+case-folding, so *deploying* matches *deploy*; drop the `porter` half to match
+whole words only.
+
+Query the index and join back to the row table on `rowid`:
+
+```bash
+dirsql query "
+SELECT n.slug,
+       bm25(notes_fts) AS score,
+       snippet(notes_fts, 0, '[', ']', '…', 8) AS excerpt
+FROM notes_fts JOIN notes AS n ON n.rowid = notes_fts.rowid
+WHERE notes_fts MATCH 'deploy'
+ORDER BY score" -c ./.dirsql.toml
+```
+
+```json
+[{"excerpt":"Rebase before you [deploy]. Keep pull requests small.","score":-1.0476190476190478e-6,"slug":"branches"},
+ {"excerpt":"…the spaghetti goes in. [Deploy] the garlic late.","score":-8.461538461538463e-7,"slug":"pasta"}]
+```
+
+**`bm25()` returns negative scores, and better matches are more negative** —
+so `ORDER BY score` ascending is best-first, with no `DESC`. `snippet()`'s
+arguments are the table, the column index, the open and close markers, the
+ellipsis, and the token budget.
+
+### Check the delete trigger
+
+A wrong or missing delete trigger fails **silently**: the index keeps rows for
+files that are gone, and they keep matching. Nothing errors.
+
+The symptom is a hit that has no row behind it, which a `LEFT JOIN` exposes:
+
+```bash
+dirsql query "
+SELECT n.slug
+FROM notes_fts LEFT JOIN notes AS n ON n.rowid = notes_fts.rowid
+WHERE notes_fts MATCH 'frost'" -c ./.dirsql.toml --persist
+```
+
+```json
+[{"slug":null}]
+```
+
+A `null` slug is a stale index entry — the base row is gone and the index did
+not hear about it. With the `notes_ad` trigger above, the deleted file's entry
+is gone too and the query returns nothing.
+
+Note the `--persist` flag: without it every run rebuilds the index from
+scratch, so a broken delete trigger cannot show itself. Deletes are visible
+against a [warm cache](./persist.md) and through the watcher, which is exactly
+where the staleness would have bitten.
+
+## Vector search (`vec0`)
+
+Two pieces beyond SQLite: the
+[`sqlite-vec`](https://github.com/asg017/sqlite-vec) extension supplies the
+`vec0` virtual table, and
+[`dirsql-plugin-embeddings`](../plugins.md#dirsql-plugin-embeddings) supplies
+`embed()`. Installing the plugin brings both — its config fragment declares
+the extension too, and the launcher
+[discovers it](../reference/cli.md#plugins):
+
+```bash
+uvx --with dirsql-plugin-embeddings dirsql query "…" -c ./.dirsql.toml
+```
+
+Loading `sqlite-vec` by hand instead (another runtime, a pinned build) is
+[Load a SQLite extension](./load-extension.md).
+
+### Find your model's dimension
+
+A `vec0` column declares a fixed width, so the template needs the number of
+components your model emits. `embed()` returns the vector as JSON text, so ask
+it:
+
+```bash
+uvx --with dirsql-plugin-embeddings dirsql query \
+  "SELECT json_array_length(embed('probe', 'minishlab/potion-retrieval-32M')) AS dims"
+```
+
+Run it once per model id you intend to use, and substitute the answer for the
+`512` in the template below. (The first call for a model downloads it — see
+[model](../plugins.md#model).) Getting it wrong is at least loud — the build fails naming both
+numbers:
+
+```
+dirsql query: failed to load config: SQLite error: Dimension mismatch for
+inserted vector for the "embedding" column. Expected 8 dimensions but received 4.
+```
+
+### The template
+
+```toml
+[[table]]
+name    = "notes"
+glob    = "notes/**/*.md"
+on-file = "python3 extract.py {path}"
+ddl     = '''
+CREATE TABLE notes (slug TEXT, title TEXT, body TEXT);
+
+-- Width must equal what the probe printed for the model id named below.
+CREATE VIRTUAL TABLE notes_vec USING vec0(embedding float[512]);
+CREATE TRIGGER notes_vi AFTER INSERT ON notes BEGIN
+  INSERT INTO notes_vec(rowid, embedding)
+    VALUES (new.rowid, embed(new.body, 'minishlab/potion-retrieval-32M'));
+END;
+CREATE TRIGGER notes_vd AFTER DELETE ON notes BEGIN
+  DELETE FROM notes_vec WHERE rowid = old.rowid;
+END;
+'''
+```
+
+The delete side is an ordinary `DELETE`, not FTS5's `'delete'` command row —
+`vec0` is a normal-looking table that way.
+
+Query it with `sqlite-vec`'s KNN form, joined back on `rowid`:
+
+```bash
+uvx --with dirsql-plugin-embeddings dirsql query "
+SELECT n.slug, v.distance
+FROM notes_vec AS v JOIN notes AS n ON n.rowid = v.rowid
+WHERE v.embedding MATCH embed('how do I cook pasta?', 'minishlab/potion-retrieval-32M')
+  AND k = 3
+ORDER BY v.distance" -c ./.dirsql.toml
+```
+
+`MATCH` plus `k = 3` is what makes this a top-k scan rather than a full one —
+`vec0` uses `k`, not `LIMIT`. `distance` is supplied by the virtual table,
+closest first.
+
+**Name the same model id in the trigger and the query.** Nothing checks that
+the two agree, and vectors from different models are not comparable. If the
+models happen to share a width the mismatch is entirely silent — rankings just
+get worse. Spelling the id out in both places, rather than leaning on the
+default in either, is the cheap defence; it also puts the model in the
+[config hash](../reference/config.md#batch-ddl), so changing it rebuilds the
+index instead of mixing old vectors with new.
+
+Rows are embedded once, at ingest, and cached on disk by content and model
+([vector cache](../plugins.md#vector-cache)). A search then costs one
+`embed()` call for the query text plus an in-process scan.
+
+::: warning Editing `ddl` wedges a persisted `vec0` cache
+Under `--persist`, editing any part of a `ddl` batch whose cache holds a
+`vec0` table currently fails with `SQLite error: no such module: vec0`, and
+keeps failing until the cache is deleted (`rm -rf <root>/.dirsql`). Tracked in
+[#1008](https://github.com/thekevinscott/dirsql/issues/1008); FTS5 is
+unaffected.
+:::
+
+## What a rebuild does
+
+`ddl` runs once, when the table is created. Editing any character of it
+changes the config hash, which drops a persisted cache and re-ingests every
+file — so a new index, a different tokenizer or a changed model id all rebuild
+from scratch rather than leaving a half-migrated index behind. The full rules
+are in [Batch `ddl`](../reference/config.md#batch-ddl).
+
+## Going further
+
+- Ranked semantic search over files with no config at all —
+  [Search documents by meaning](./search-by-meaning.md).
+- Why indexes belong to declared tables and not to path-tables —
+  [how `dirsql` thinks](../explanation.md).

docs/reference/path-tables.md

@@ -190,7 +190,13 @@ That is the right trade for a hundreds-of-files, run-it-once question. When the
 same tree is queried repeatedly, or is large, declare a
 [table](/reference/config) for it instead — a declared table is indexed on
 build, kept fresh by the watcher, and (with `--persist`) survives restarts, so
-its rows are read from SQLite rather than re-walked each time.
+its rows are read from SQLite rather than re-walked each time. It is also the
+only place a keyword or vector index can live
+([templates](/howto/search-indexes)).
+
+Neither kind is the better one: each buys what the other gives up, and the
+choice is stated as one trade in
+[how `dirsql` thinks](/explanation#two-table-kinds-opposite-trade-offs).
 
 ## Skip rules