Module store 

SQLite side of the store: file bookkeeping and chunk payloads.

The vector index only ever holds (chunk id, quantized vector). Everything a caller actually wants back — the text, the path, the line span — lives here and is looked up by chunk id after the search returns.

Modules

class

The classes a not-indexed row falls into (2.3), and the dummy list.

Structs

Acceptance

A person’s decision about one file.

ChunkRow

A chunk row joined with its file’s path — what a search hit resolves to.

Corpus

Everything the Inside the index page shows about one store.

EdgeEnd

One end of a traversal: the symbol reached, and the edge that reached it.

Facets

What a listing spans: the distinct file extensions in it, and the total number of lines across every file.

FileKind

One kind of file on the prose panel: how many there are and, for a kind with a unit, how many of that unit they hold.

FileLines

FileRow

One row per indexed file: path, bytes, chunks, the last line any chunk of it covers, and when it was last indexed.

Filled

What a reader found for one row: the usage, or only why there is none.

ImageRow

An indexed image: its file’s path and the pixel size.

Incoming

One edge into a name, with the definitions of that name it resolves to.

KindCount

LanguageCoverage

What the graph covers, per language, read from the rows themselves.

LanguageLines

LedgerFact

One ledger row, every column a figure is made of.

Misses

The two figures for what semlith did not answer.

MonthCount

NamedEdge

One settled dependency edge, by name, for the community pass.

NewRetrieval

One retrieval, before it has a timestamp or a place in the chain.

PastSymbol

A definition a re-index replaced.

ReadFile

A file the store read at or after since, and what it held before.

Refused

One file or folder that was not indexed, and why.

ReportRow

One ledger row as the reports count it (0.42.0).

Retrieval

One recorded retrieval.

RetrievalRow

One recorded retrieval, as the access report attaches them.

RowUsage

A row’s usage columns as they are stored. Every figure is optional: a client that logs the model and not the tokens gets the model.

Savings

What the ledger adds up to.

SessionRow

One session’s whole ledger, as the per-session table reads it.

Stamp

What a file looked like when it was read for the store, and the one rule every surface judges “edited since it was indexed” by.

SymbolRow

A symbol row joined with its file’s path — what a graph answer resolves to.

SyncRow

One ledger row as ledger sync sends it: what was retrieved, never the query text, which this struct does not carry.

Unfilled

A ledger row whose usage has not been looked for, as the readers in crate::usage need it to find the model request that made the call.

Unresolved

An edge whose target this store has no definition for.

Writing

Permission to write, given back when it goes out of scope.

Enums

FileSort

The column a Files listing is ordered by.

LedgerScope

Which rows of the ledger a count is about.

Constants

AGENT_RETRIEVAL

The rows that are an agent’s retrievals, as SQL: what ledger::figures counts by default. Not a refund (raw-read), not the chain’s own bookkeeping (a note, a refused write, an accept or revoke), and not a person at the portal or the CLI. Every per-store reader below takes it, so a store’s card and its report say what the Ledger page says.

BASIS_BYTES

What the whole-file side of a row was counted from: the file’s bytes.

BASIS_TEXT

What the whole-file side of a row was counted from: the text a reader extracted from it, counted the way the excerpt was.

CODE_CONTEXT

The first format whose code chunks embed the definition they sit inside.

DEFINITION_CHUNKS

The first format whose Markdown chunks are cut at their headings.

FIRST_IMPORT_KEY

Meta key present while a store’s first import is under way (#210).

FORMAT_KEY

FORMAT_VERSION

The store layout this binary understands.

NOTE_TOOL

The tool of a ledger note: a row the chain carries that is not a retrieval, and no total counts.

REFUSED_TOOL

The tool of a call refused because the agent key is read-only; query names the tool. Not a retrieval, so no total counts it either.

SHARDED_FORMAT

The first format that keeps its vectors in shards.

TURBOVEC_V8

The first format whose vector files may be turbovec’s v8.

USAGE_COLUMNS

The usage columns, from column at of a row onwards, in the order USAGE_COLUMNS lists them. None until a reader has looked.

Functions

accept

acceptance

acceptances

Every acceptance, for counts and for the folders the walk must not prune.

all_hashes

Every stored path’s recorded hash, read once for a whole index pass so its prepare stage can tell an unchanged file without the database.

all_paths

bounded

Run work over conns, interrupting their queries once limit passes.

builtin_targets

How many distinct call and import targets with no definition here are the language’s own builtins or standard library, and so are not in unresolved_targets.

busiest_names

The names with the most edges in or out, busiest first, each counted once on its own index. For the graph page’s choice of where to open: a name with seven hundred definitions is one candidate, not seven hundred.

chunk

chunk_count

Every chunk row, durable or not: what the sidecar can hold a record for.

chunk_ids_of

Drop a file and its chunks, returning the chunk ids so the caller can evict them from the vector index too. The ids of a file’s chunks.

chunking

Which chunking rule this store’s chunks were cut by, in the words stats and the portal print.

chunks_by_month

Chunks by the month their file was indexed, oldest first.

chunks_overlapping

Every chunk of one file whose lines overlap start..=end, in file order.

community_edges

Every calls and imports edge whose target this store defines, up to limit, with the total so a reader is told when the list was cut.

corpus

Measure one store, for the Inside the index page.

coverage_by_language

Per-language coverage for every language the store holds.

delete_file

delete_file_unretired

delete_file without keeping the file’s definitions as history: for a stopped run’s undo, whose files were never the store’s to remember.

delete_graph

Remove one file’s symbols, and with them the edges that leave them.

drop_insert_trigger

Drop the insert trigger an older semlith’s schema puts back when it opens this store, inside the writer’s transaction so it cannot come back before the commit. With it in place every chunk would be indexed twice.

durable_chunks

Chunks whose file has a committed hash — the ones whose vectors are on disk.

edge_counts

How many edges touch name, and how many of them are certain, counted on the indexes alone: (resolved, total).

edges_in

What points at name: callers, importers, references in.

edges_in_first

At most limit of edges_in, the extractor’s certain edges first.

edges_in_resolved

What points at name, each edge resolved against the name’s definitions the way edges_out resolves it from the other end.

edges_out

What the symbols named name point at: callees, imports, references out.

edges_out_short_of_hubs

edges_out without the edges whose target has more than hub definitions, dropped before their candidates are joined.

edges_out_to

edges_out restricted to edges whose target is one of targets, settled exactly as edges_out settles them.

end_first_import

The first import ran to its end: later passes write as any other does.

excerpt_tokens_by_tool

(tool, excerpt_tokens) for every ledger row that names its tool, the newest limit of them: what the Agents page’s typical answer sizes are the median of.

file_count

How many indexed files match groups.

file_facets

file_hash

Content hash recorded for path, if we have indexed it before.

file_rows

The first take matching files in the requested order.

file_stamps

(bytes, indexed_at) for each of paths the store knows.

files_ending_with

Indexed files whose path ends with suffix, most specific first.

files_indexed_since

Files this store read since since, newest first, with when it read them.

files_read_since

Every file read at or after since, newest first, with no cap — the change brief’s “Files re-read” was 200 per store summed (R3).

fill_usage

Write a row’s usage. Every row sharing its query_id is the same call, written once per store that answered, so all of them are filled.

filtered_chunk_ids

Ids of every chunk whose file matches groups.

filtered_image_ids

The image ids whose file matches groups, for a filtered search.

filtered_paths

Indexed paths whose file matches groups, in path order.

first_import

Whether this pass is part of a store’s first import, starting one when the store is empty.

format

Which layout this store’s vectors are in.

free_bytes

Pages SQLite holds free inside the file, in bytes: what a vacuum gives back before anything else is dropped.

get_meta

graph_input

A file’s id and its chunks’ line spans, for writing its graph again.

graph_stats

(symbols, edges) — the graph’s size, for stats and the portal.

grep_cost

The total bytes of every file a grep for name would have to read.

heal_stamps

Bring an unchanged file’s bytes and indexed_at up to what is on disk, when the file’s size or mtime has moved past the row.

held_text

The text the store holds for path, stitched by line from its chunks.

history_bytes_before

Roughly what the definitions retired before before take: their text plus a row’s fixed cost. An estimate for stats, not an accounting — SQLite packs rows into pages and the count cannot see that.

history_count_before

How many definitions were retired before before.

hold_automerge

Stop FTS5 merging segments while a first import writes (#210).

holds_path

Whether this store has path indexed.

image

One image row, joined with its file’s path.

image_count

How many images this store holds.

image_ids

Every image id, for the ids a filtered search may consider.

image_ids_of

The image ids belonging to path, before its row is deleted.

image_rows

Every indexed image, for the decisions table’s “read as image” rows.

insert_chunk

insert_edge

insert_file

insert_image

insert_symbol

is_interrupted

Whether an error is a query bounded interrupted.

is_pinned

Whether this connection is holding a run’s snapshot.

keyword_search

keyword_search_within

The keyword list inside a set of chunk ids, for a filtered search that has the set already.

last_write

The newest indexed_at this store holds, or None when it holds nothing.

ledger_break

Re-walk the chain and return the id of the first row that does not verify.

ledger_clients

How many retrievals each client made, most first.

ledger_facts

Every ledger row with since <= at < until, oldest first.

ledger_head

The hash of the newest ledger row: what --verify checks the chain to.

ledger_keys

The distinct retrievals this store holds, as keys that can be unioned across stores.

ledger_latency

Warm query latency as the ledger recorded it: p50 and p95, in microseconds.

ledger_misses

What the ledger says about the questions semlith did not answer.

ledger_note

Re-anchor the chain after a row that does not verify, by appending a note row onto the current head (0.35.0’s Re-verify with repair).

ledger_repair

Append a note for every unacknowledged break, first to last, and say which rows they were. Bounded, so a chain broken at every row is not a loop.

ledger_retrievals

Every retrieval since since, newest first.

ledger_rows

How many rows the chain holds, notes included: what a verify walked.

ledger_savings

The one savings figure, with the denominators that make it readable.

ledger_savings_since

The savings, over retrievals recorded at or after since.

ledger_sessions

Every session this store recorded, newest last-seen first.

ledger_totals

(queries, clients, excerpt tokens, whole-file tokens) over the whole/// (queries, clients, excerpt tokens, whole-file tokens) over the whole ledger — what the Ledger page shows.

ledger_zero_hit_queries

Queries that found nothing, commonest first.

ledger_zero_hit_queries_since

The questions that found nothing, at or after since.

mark_synced

Record that these rows reached the org’s ledger.

mark_vector_format

Move a format-4 store to TURBOVEC_V8 before this binary writes to it.

matching_files

How many indexed files groups selects.

mtime_ns

A file’s modification time in nanoseconds since the epoch.

names_with_several_definitions

How many names this store defines more than once.

newest_chunk

The highest chunk id, or 0 for none. One step down the rowid b-tree.

open

Open a store’s database, laying down or upgrading its schema first.

owner_of

The type or module a definition’s qualified name puts it in: Fleet for Fleet::search_preferring, store for store::edges_out.

path_bounds

The first and last indexed path in sort order. Their common directory is the common directory of every path, which is the root a report prints paths against when the registry names none (a store opened by --store).

pending_among

Of ids, the chunks whose file is still being embedded, with their text: what a search mid-run embeds on the query path.

pending_share

The share of the store’s chunks still being embedded, or None when nothing is: one indexed probe on every search, and the counts only when the probe finds a run in progress.

pin_snapshot

Hold this reader at the store as it is now, until unpin_snapshot.

prune_history

Drop the definitions retired before before (unix seconds). The retention is the caller’s; see compact.

read_only

Make this connection refuse writes until a writer says otherwise.

readable

(files, chunks, indexed bytes) Whether this store can still be read, as cheaply as that can be asked.

reclaim

Give the database’s dead pages back to the filesystem.

record_counted

record_retrieval, saying what the whole-file side was counted from.

record_legacy_retrieval

Append a row in the shape 0.14.0 wrote, for the test that proves a store holding both kinds still verifies.

record_retrieval

Append one retrieval, chained to the row before it.

record_v1_retrieval

Append a row the way 0.15.0 to 0.41.0 wrote it — chain v1, no basis, no head — for the test that proves a store holding v1 and v2 rows verifies.

refusal

One not-indexed row, by exact path.

refusals

Every not-indexed row, reviewable first, then by class and path.

refuse

Record that path was not indexed. The first time it was seen is kept.

release_automerge

Put back the automerge setting hold_automerge recorded, if any.

report_rows_since

Every ledger row recorded at or after since, oldest first.

retrieval_at

When a ledger row was written, by id.

retrievals

The most recent limit retrievals, newest first.

revoke

salt

The per-store salt fingerprints are keyed with, made on first need.

salt_if_any

The salt, when a store has one. A read-only caller cannot make one, and a store without one has no acceptances to compare against.

secure_delete_fts

Turn on FTS5’s secure delete for chunks_fts, if it is not on: a removed row’s terms then leave the index’s segments at once rather than at their next merge, so a file refused for a secret leaves no searchable term behind. A setting the index keeps; older binaries read it unchanged.

set_file_graph

Record what the parser made of a file: parsed, timeout, or none for a language semlith carries no grammar for.

set_file_units

Record a file’s units (see chunk::COUNTED), by row id or, for a file whose row is already there, by path.

set_meta

set_read_mtime

Record the modification time a file’s read saw. See Stamp::fresh.

set_units_by_path

stats

symbols_by_names

Symbols defined in files matching groups whose name is one of names.

symbols_in_files

Every symbol defined in each of paths.

symbols_named

Every definition of name, across the whole store.

symbols_named_counted

symbols_named with code first — definitions before the headings, keys and selectors that share the name — and how many there are in all.

symbols_past_count

How many definitions this store has retired, over all names.

symbols_past_named

What a symbol used to be, newest first.

symbols_scoped

Symbols to draw for a scope, newest-defined first within each file.

text_of

Everything this store holds for one file, its chunks in order, joined.

uncounted

Stored files of a counted format whose units were never recorded — written by a binary older than 0.34.0. The index pass reads each one again for its count alone.

unfilled

Rows written by an AI client since since (unix seconds) with no usage yet, oldest first. cli and portal rows have no model behind them, and neither has a raw-read row, which the steering hook writes for a read the agent made without semlith.

unpin_snapshot

End pin_snapshot’s read transaction: the reader sees the run’s commit.

unrefuse

Forget that path was not indexed: it has been, or it is gone.

unresolved_out

Targets of name’s edges that the store holds no definition for.

unresolved_targets

Call targets no definition in this store satisfies, commonest first, and how many distinct ones there are.

unsynced

Rows written at or after since and not yet sent, oldest first, at most limit.

unsynced_count

How many rows since since are still to send.

usage_at

usage_rows

Every row with usage, keyed as ledger_keys keys it, for totals per client and per model that count a cross-store call once.

where_defined

Where each of these names is defined, for the hub lines under a community.

write_extraction

Write one file’s extraction: its symbols, then its edges by source name.

Type Aliases

ChunkSpan

A chunk’s first line, last line and id.

SymbolSpan

One definition’s span and identity: (start, end, name, kind).