ChatStreamingView built an IN filter from every permitted document id
for the "chat over all documents" case, which exceeds the vector
store's SQLite bound-parameter safety limit on installs with more
than ~32700 documents, silently returning no context. For a user who
can see every document (an active superuser), that filter never
narrows anything, so skip it and let the retriever search the whole
index instead.
AI Suggestions previously invented near-duplicate metadata because the classification
prompt had no knowledge of the installation's own taxonomy. This surfaces
a small, ranked, permission-filtered set of existing tags/document
types/correspondents/storage paths - drawn from the document's RAG
neighbors plus its own already-assigned metadata - so the model prefers
reusing what already exists.
The LLM response schema now returns existing_ids (IDs of reused
candidates) separately from new_names (genuinely new suggestions).
Only new_names goes through localization and fuzzy name-matching;
existing_ids is resolved deterministically and never touched by the
localization pass, so exact matches can no longer be silently
corrupted by translation.
In tracemalloc based profiling, not materializing the whole Document list
reduced memory to approximately 20% of the baseline, with a peak memory
that scaled with the library size. Now, the lazt queryset is used and only
the needed pk value is actually contributing to memory
get_context_for_document() always materialized every document id a user
can see into a Python list, even for a superuser (or no user at all),
passing it through as a SQL IN filter. For a superuser, that's the whole
library:
- Past ~32,763 documents, this crashes outright:
sqlite3.OperationalError: too many SQL variables (SQLite's
SQLITE_MAX_VARIABLE_NUMBER is 32766 by default, and the query already
binds embedding + k + a NE self-exclusion clause alongside the ids).
- Below that cliff, vec0's IN-list evaluation is a nested loop (strncmp
per row per allowed id), so it's quadratic in library size for no
reason -- the filter was never going to exclude anything.
get_objects_for_user_owner_aware() already returns every Document for a
superuser (guardian's own with_superuser shortcut), so skipping straight
to document_ids=None changes nothing about which documents are
considered, only how we get there.
Also:
- Drop the pointless sorted() in _document_id_filters(): a SQL IN clause
doesn't care about order, so it was pure overhead on every call.
- Add a hard guard in _build_where() so a future regression (or a
legitimately huge permission-restricted user) fails closed -- no rows,
a logged warning -- instead of a cryptic OperationalError. Since this
filter scopes document access, failing closed rather than skipping the
filter is the only safe way to handle an oversized list.
First item from VECTOR_STORE_PERF_BACKLOG.md (an audit done alongside
perf/13314-vecstore-point-delete, deferred to its own branch since that
one was already large).
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* Fix: make migration-check result tri-state to avoid deferred-vs-current ambiguity
* Test: cover update_llm_index()'s migration-check-deferred branch
* Refactor: rename COMPACT_BATCH_SIZE to BATCH_SIZE, no longer compact()-specific
* Feature: schema v2 -- document_chunks/document_meta side tables, document_id INTEGER, point-delete
Rewrites the sqlite-vec vector store's on-disk schema: document_id becomes
an INTEGER vec0 metadata column (was TEXT), modified moves out of vec0 into
a new document_meta side table, and a new document_chunks side table gives
O(1) per-document chunk lookup for delete/upsert instead of a full vec0
scan. compact() now streams document_chunks and document_meta across the
file-swap rebuild too (previously document_meta would have gone silently
empty after the first compaction). drop_table() clears both side tables.
Adds the single frozen m0001_v1_to_v2 migration, converting a real,
historically-shaped v1 store (the shape shipped since v3.0.0) into the v2
shape in one streaming pass, with its own hardcoded DDL rather than
delegating to any "current schema" helper.
SCHEMA_VERSION bumps 1 -> 2.
* Fix: strengthen two vacuous Task 4 regression tests
test_migration_never_delegates_to_current_schema_helpers never actually ran
the migration (missing check_and_run_migrations() call) and its source-text
assertion was tautological (the "or DROP TABLE in source" clause was always
true). Now runs the real migration and asserts spy call counts instead:
DocumentChunksTable.create/DocumentMetaTable.create are each called exactly
3 times (construction, rebuild temp file, post-swap reopen -- all via
_open_connection, never from inside apply()), and _create_vec_table is
never called from the migration path.
test_drop_table_clears_modified_times asserted via get_modified_times(),
which short-circuits on table_exists() -- checking only the vec0 table that
drop_table() drops first -- so the assertion held even if
DocumentMetaTable.delete_all() were never called. Now asserts directly
against document_meta and document_chunks row counts.
* Perf: dedupe table_exists() lookups, atomic insert counter, fewer connections in update_llm_index()
* Fix: guard compact() against unmigrated stores, apply final-review cleanups
compact() had no migration guard: on a v1-schema store, document_chunks
reads 0 (freshly created empty) while total_inserts reflects the real
cumulative count, so the bloat check nearly always rebuilt -- silently
losing document_meta (copy_all reads from the empty v1 table) while
schema_version copied across unchanged, leaving the store permanently
unmigratable. compact() now calls has_pending_migration() and no-ops with
a warning instead.
Also folds in five minor final-review findings: drop _rebuild_into's
unused int return, hoist test-local imports to module level in
test_vector_store.py, note in TestMigrations' docstring that its fake
structural migrations only exercise dispatch (not full schema
correctness), restore the comment explaining why _row() requires
document_id, and tighten increment_total_inserts' docstring to not imply
general concurrency safety beyond its single atomic statement.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
build_llm_index_text queried Note and CustomFieldInstance (plus its
field FK) per document, uncovered by the earlier correspondent/type/
storage_path/tags prefetch fix. Add notes and custom_fields__field to
the rebuild and scoped document querysets, and read from doc.notes.all()
instead of a fresh Note.objects.filter() so the prefetch is actually used.
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* Perf: scope bulk_update_documents' LLM index refresh to the edited document ids instead of the whole library
* Perf: select_related/prefetch_related for the scoped LLM index batch
* Perf: select_related/prefetch_related for the full LLM index rebuild path
* Fix: address Copilot review findings on LLM index scoping
query_similar_documents() never excluded the querying document itself, so a document could appear in its own "similar documents" context, duplicating its content into the AI-suggestions prompt and inflating prompt size/tokens unnecessarily. Add NE filter support to the vector store and exclude the source document's id at query time.
* Chore(beta): add sqlite-vec 0.1.9 dependency
Pinned exactly: the 0.1.9 wheels carry no baked SIMD flags (safe on
pre-AVX2 CPUs, the point of this migration); the 0.1.10 alphas bake
-mavx and would reintroduce the #12970 crash class.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Test(beta): port vector store tests to sqlite-vec backend
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Enhancement(beta): switch AI vector store from LanceDB to sqlite-vec
Fixes the non-AVX2 SIGILL class (#12970) at the root: lancedb is no
longer imported. sqlite-vec 0.1.9 wheels carry no baked SIMD, vec0
metadata columns give parameterized EQ/IN filtering, WAL preserves the
lock-free-reader model, and compact() rebuilds the table because vec0
DELETEs never reclaim space.
Implementation notes vs. the Task 3A draft:
- compact() uses a file-swap approach (new db file + Path.replace) rather
than ALTER TABLE RENAME, which does not cascade to shadow tables in
sqlite-vec 0.1.9 (upstream limitation).
- Bloat is tracked via a cumulative total_inserts counter in index_meta
because the _rowids shadow table does not accumulate deleted rows in
0.1.9 (contrary to the design doc assumption from #54).
- None distances from the zero-vector cosine edge case are mapped to
similarity 0.0 rather than raising TypeError.
- Test suite updated accordingly: _bloat_ratio reads index_meta instead
of _rowids; seed collision in force-compact test fixed (seed=100.0).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Enhancement(beta): wire indexing pipeline to the sqlite-vec store
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Enhancement(beta): move filename/storage path/ASN to node metadata
Same treatment as title/tags/correspondent in #12944: excluded from
the embedded text, visible to the LLM via metadata prepend. Changes
embedded text for every document, so it ships inside the sqlite-vec
transition, whose forced rebuild re-embeds everything anyway.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Test(beta): cover legacy LanceDB index cleanup and forced rebuild
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore(beta): drop lancedb dependency
Fixes#12970: the package whose wheels SIGILL on non-AVX2 CPUs is no
longer installed at all.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore(beta): partial pyrefly cleanup on sqlite-vec vector store
- Add MetadataFilter import and isinstance guard in _build_where()
- Add query_embedding None guard in query()
- Fix dict.get() type-checker ambiguity in get_configured_model_name()
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore(beta): drop automatic LanceDB index cleanup on startup
Leave legacy Lance directory removal to the user rather than deleting it
automatically on first run. Beta policy: user is expected to do a clean
re-embed anyway; no need for the system to silently delete their data.
Remove _cleanup_legacy_lance_index(), the forced-rebuild path that called
it, and the associated tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore(beta): ruff format pass on sqlite-vec AI files
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Removes the benchmarking file
* Try to resolve or silence some semgrep. But we're using SQL here, not an ORM and we control the inputs, not users
* Enhancement(beta): add schema migration machinery to sqlite-vec vector store
Adds versioned schema migration support modelled after PR #12968's LanceDB
approach, adapted for sqlite-vec's file-swap compaction pattern.
- SCHEMA_VERSION = 1 written to index_meta at table creation and preserved
through compact()
- Migration dataclass with from_version, to_version, kind ("structural" or
"re-embed"), description, and an optional apply(src, dst, dim) callable
- MIGRATIONS registry (empty at v1 baseline); add entries and bump
SCHEMA_VERSION when the schema changes
- check_and_run_migrations(): structural migrations run via the same
file-swap as compact() (no re-embed); re-embed migrations return True
so the caller forces a full rebuild
- update_llm_index() calls check_and_run_migrations() under the write lock
before any indexing work
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore(beta): deduplicate vector store internals via helper methods
Extract three helpers to remove copy-paste between compact() and
_run_structural_migration():
- _meta_set_on(conn, key, value): static upsert into any connection's
index_meta; _meta_set() now delegates to it
- _create_vec_table(conn, dim): CREATE VIRTUAL TABLE DDL (carries the
nosemgrep annotation)
- _swap_in_compact(compact_path, db_path): close/replace/reconnect
sequence used by both file-swap callers
Also normalises compact() error-path cleanup to unlink(missing_ok=True).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Adds equality test and no covers some defensive error handling stuff
* Ensures an embed migration stops the migration chain, just in case
* Silence one kind right but not really semgrep
* Trims dead assignment
* Fix(beta): address Copilot review on sqlite-vec vector store
Three findings from the PR review:
- compact() failure cleanup now unlinks the temporary .compact-wal and
.compact-shm files, matching _run_structural_migration(); previously
only the main .compact file was removed.
- _build_where() fails closed (1 = 0) when filters are requested but none
translate, instead of emitting "()" which is invalid SQL; filters scope
document access, so an empty translation must match no rows.
- Drop the unused table_name constructor parameter (all SQL hardcodes
DEFAULT_TABLE_NAME) and its callers in indexing.py.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Enhancement(beta): guard sqlite-vec compaction swap against concurrent readers
The compaction/migration file swap replaces the database via os.replace,
but the -wal/-shm files are keyed by path, not inode. A reader holding an
open connection across the swap leaves the old WAL aliased onto the new
file; a subsequent write then corrupts the database (reproduced via
PRAGMA integrity_check).
Add a cross-process read/write lock (filelock.ReadWriteLock) over the
index:
- read_store() holds it shared for the whole connection lifetime (and
closes the connection on exit); concurrent readers do not block.
- compaction and the migration check run under an exclusive lock that
drains readers, and skip with an info log on Timeout (maintenance op,
retries next run).
- Normal writes are untouched: WAL gives reader/writer concurrency and
LLM_INDEX_LOCK still serializes writers, so they never block readers.
load_or_build_index() now takes the store from the caller's read_store()
so the lock and connection span the whole retrieval; chat holds it across
the streamed response. Two new settings: LLM_INDEX_RWLOCK and
LLM_INDEX_COMPACTION_LOCK_TIMEOUT.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Ensures the store alays cleans up SQLite connections for any operations, even on errors
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Move LLM index lock outside index dir and skip per-doc tasks on bulk update
Two concurrency bugs from #12893:
[P1] Lock file lived inside LLM_INDEX_DIR. A rebuild calls
shutil.rmtree(LLM_INDEX_DIR), deleting the lock while a worker still
held it. A second worker then acquired a fresh lock on the new path and
ran concurrently, defeating serialisation. Move the lock to
DATA_DIR/locks/llm_index.lock (a new settings constant LLM_INDEX_LOCK)
so rmtree cannot touch it. The locks/ dir is created at settings load
time, matching the existing pattern for LOGGING_DIR.
[P2] document_updated was connected to add_or_update_document_in_llm_index
in apps.py. bulk_update_documents() emits document_updated for every
document in the batch, queuing N per-document LLM tasks, and then also
calls update_llm_index(rebuild=False) once at the end. Pass
skip_ai_index=True when sending document_updated from the bulk path so
the handler skips the per-document enqueue; the existing batch call at
the end of bulk_update_documents is the only LLM update for that path.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: ghost vectors leave KeyError-prone nodes_dict entries after deletion
docstore.delete_document() removes a node from the docstore but leaves its
entry in index_struct.nodes_dict (the FAISS positional-id to node-UUID map).
A subsequent similarity query resolves the ghost position to the deleted UUID,
finds nothing in fetched_nodes_by_id, and raises KeyError inside
_insert_fetched_nodes_into_query_result.
Purge stale nodes_dict entries after each docstore deletion and re-sync the
mutated index_struct into the kvstore so persist() writes the updated mapping.
Dead FAISS vectors remain in the flat index until the next full rebuild
(IndexFlatL2 is append-only); add a try/except KeyError around
retriever.retrieve() as a defensive fallback for any residual ghost positions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: acquire index lock in query_similar_documents
query_similar_documents() loaded the index and ran the FAISS retriever
without holding the file lock. All write paths (update_llm_index,
llm_index_add_or_update_document, llm_index_remove_document) hold
FileLock(_index_lock_path()), so a concurrent rebuild calling
shutil.rmtree(LLM_INDEX_DIR) while a read is mid-load produces an IOError
or corrupt partial state.
Wrap the load_or_build_index() call and all subsequent retriever work inside
FileLock. The early-return guards (vector_store_file_exists check, empty
allowed_document_ids) remain outside the lock; the DB query for the final
result set also stays outside.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: skip LLM index enqueue on document_updated during version addition
When a document is consumed as a new version of an existing document, the
consumer fires document_consumption_finished (which triggers
add_or_update_document_in_llm_index) and then document_updated for the root
document. Both signals are connected to the same handler, so the root document
was enqueued for LLM indexing twice per version-addition event.
Pass skip_ai_index=True on the consumer's version-addition document_updated
send so the handler's existing guard suppresses the duplicate enqueue.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Test: bulk_update_documents must not enqueue per-doc LLM tasks
With AI enabled, bulk_update_documents() sends document_updated for every
document in the batch. The skip_ai_index=True kwarg (added in the P2 fix)
prevents add_or_update_document_in_llm_index from enqueuing a per-document
task for each one. Only the single update_llm_index call at the end should run.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Debug level log sure
* Update src/paperless_ai/indexing.py
Co-authored-by: shamoon <4887959+shamoon@users.noreply.github.com>
* Apply suggestion from @shamoon
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: shamoon <4887959+shamoon@users.noreply.github.com>
* Fix: Remove all nodes for multi-chunk documents in update_llm_index incremental path
The existing_nodes dict comprehension keyed on document_id silently dropped all
but the last node per document, so only that one node was deleted when a
modified document was re-indexed, leaving all other chunks as ghost vectors in
the FAISS index. Switch to a defaultdict(list) that collects every node per
document_id, then iterate and delete all of them before inserting fresh nodes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Wire document_updated signal to LLM index update handler
Connect document_updated to add_or_update_document_in_llm_index in
DocumentsConfig.ready() so REST API edits (PATCH /api/documents/{id}/)
enqueue an LLM vector store update, matching the existing
document_consumption_finished behavior.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Add file lock around FAISS index mutations to prevent concurrent write corruption
Two concurrent Celery workers calling llm_index_add_or_update_document or
llm_index_remove_document each loaded the same on-disk index independently,
made their own change, and the last writer silently overwrote the first's
update. Wrap both functions and the rebuild/persist body of update_llm_index
in a filelock.FileLock keyed on LLM_INDEX_DIR/index.lock. Add a TOCTOU
comment on queue_llm_index_update_if_needed explaining the residual risk
(duplicate rebuild tasks are wasteful but not corrupting because the lock
serialises the actual write).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Apply _normalize() in extract_unmatched_names to prevent duplicate suggestions
extract_unmatched_names was using .lower() while _match_names_to_queryset
uses _normalize() (which also strips punctuation). A name like "J. Smith"
matched to existing correspondent "J Smith" would still appear in the
unmatched list, causing duplicate object creation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Skip LLM index update gracefully when document has no indexable content
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Persist empty index when all documents are deleted to clear stale FAISS vectors
The early-return guard in update_llm_index fired before persist() when no
documents existed, leaving a stale on-disk FAISS index that returned phantom
hits for deleted document IDs. Now the guard only returns early for the
incremental (rebuild=False) path when no index exists on disk; the rebuild
path always continues through to persist(), producing an empty clean index.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore: Simplify incremental index update — use docs.values() and deduplicate node extend
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Validate and limit chat question input in ChatStreamingView
Add max_length=4000 to ChatStreamingSerializer.q and replace the bare
request.data["q"] read with proper serializer.is_valid(raise_exception=True)
so oversized or missing questions are rejected with HTTP 400 before
reaching the LLM.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: Add defensive prompt framing to mark document content as untrusted
* Also adds a system prompt which is treated higher that this is untrusted stuff
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
stumpylog
shamoon
Moritz StücklerandClaude Sonnet 4.6