Every chat question and AI suggestion request builds the embedding model, and sentence-transformers
checks the Hub for a newer revision each time. With the model already cached that check cost about
2 seconds per request (0.2 seconds without it), and it fails outright when the Hub is unreachable.
Load from the local cache first and only download when the model is missing. A full rebuild still
checks the Hub, so a new model release is picked up together with re-embedding every document
instead of silently mixing revisions.
The vec0 default of 1024 rows per vector chunk makes each chunk a
multi-megabyte blob. On spinning disks, indexing a document took about
2.5 seconds and re-indexing one about 13 seconds. With 32 rows per
chunk the same operations take about 35 and 180 milliseconds, with the
write-ahead log a third of the size and no cost in query speed or file
size.
Pair that with 64 KiB SQLite pages, which made cold scans about 35%
faster and warm ones about 45% faster at no write cost once the chunks
are small. The page size is set before the file enters WAL mode, so it
applies to new files.
Document chat held the shared index read lock for the whole request,
including the LLM stream, which can take minutes. A compaction or
migration swap waiting for exclusive access could time out behind it
and be skipped until the next run.
Now that the answer is synthesized from the nodes already retrieved,
nothing after the vector search touches the store. Hold the lock only
while loading the index and searching, and release it before synthesis.
Document chat searched the vector store twice per question: once
explicitly to build the reference list, and again inside
RetrieverQueryEngine.query(), which retrieves before synthesizing. That
meant two query-embedding calls and two full vector searches each time.
The engine added nothing beyond the retriever call (no node
postprocessors are configured), so drop it and synthesize directly from
the nodes already retrieved. The streaming response is unchanged.
* Fix: use a root document's effective content for the LLM index and prompts
The search index and the document views use the newest version's content as a
root document's content, but everything the LLM side reads used the root's own
content: the text embedded in the LLM index, the content in the classification
prompt, the context blocks from similar documents, and the text that is
searched for similar documents. After a document was replaced by a new
version the LLM kept answering from the old text.
Read get_effective_content() in those four places. The index update annotates
the content in the query like the search index does, and the similar-document
lookup annotates it for the documents it fetches, so no query is made per
document. Entries already in the LLM index keep the old text until the
document next changes or the index is rebuilt.
* Chore: test the LLM query text and bulk annotation use effective content
* Removes the query count testing. It's flakey from local to CI and workers, doesn't really add much imo
* Fix: index document versions as their root document
The search index and the LLM index hold root documents only, a root being
indexed with its newest version's content. Every write path therefore had to
remember to hand them the root. Several did not.
WriteBatch.add_or_update, WriteBatch.add_or_update_ids and
llm_index_add_or_update_document now resolve a version to its root themselves.
The reindex command and update_llm_index only walk root documents, and an
incremental LLM update is scoped to the roots of the given ids through
versioning.root_document_ids, which add_or_update_ids shares. more_like_id
returns the root's id for a version. The reprocess task no longer picks the root
for the indexes and only still clears the caches of both documents.
* Fix: rebuild old search indexes and look versions up as their root for similar documents
An index built before versions were indexed as their root can still hold
version entries under their own id and metadata, and only a rebuild clears
them. Bump the schema version so the index is rebuilt on upgrade.
more_like_this_ids looked the given id up in the index as is, so asking for
documents like a version, as the AI suggestions endpoint can, silently found
nothing. It now looks the root up instead.
* Fix: authorize document versions by their root and speed up permission id sets
permitted_document_ids judged a version by its own owner and grants, so a
version whose owner had drifted from its root's was visible to the wrong
people and hidden from the right ones. Callers patched this individually by
mapping each document to its root first. The query itself was also slow on
MariaDB: the guardian grants were a UNION cast to integers and tested with
IN inside an OR with the owner checks, which MariaDB cannot materialize, so it
re-scans the user's grants for every document. At 20k documents that took
seconds for a user with a couple of hundred grants.
permitted_object_ids now casts the row key to a string, as guardian stores
object_pk, and tests it against a single uncorrelated UNION ALL of the user's
and groups' grants. No integer cast is needed and the user's groups are
matched with an IN subquery rather than a join through the membership table.
Postgres and SQLite build the grant set once, and MariaDB probes guardian's
unique indexes per row, which is cheap. A correlated EXISTS per grant also
fixed MariaDB but was up to 3x slower than before on Postgres and SQLite.
permitted_object_ids also takes an optional parent_field naming a
self-referencing foreign key whose target authorizes the row, and
permitted_document_ids passes root_document, so a version is visible exactly
when its root is. The helper that mapped documents to their roots at the call
sites is no longer needed, so the email, selection data, share link bundle,
trash, bulk download and bulk edit checks use the id set directly.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
TantivyBackend(path=None) built an in-memory index that is not used by production ever used. The backend now requires a path, the open, write-batch and rebuild branches are gone, and the shared backend fixture and the fulltext similar-documents fixture build a real index under the per-test directory layout.
The model factories lived in the documents test package, but three other
apps needed them. The AI, mail and testing suites all reached across an app
boundary to import from documents.tests.factories, which made a private test
package into a shared dependency.
The factories now live in the shared testing package, where cross-app use is
the intended use.
This brings users without an embedding backend configured to closer
parity with those who do. Reuse the search backend to locate similar
documents and use them to provide the LLM with the better suggestion pool
to draw from
* Fix: skip vector store document id filter for unrestricted chat users
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.
* Minor improvements from a Claude review
* When a user is unrestricted chatting, still exclude trashed documents using a 'NOT IN' SQL statement. Wire that up where we need it
* Update src/paperless_ai/chat.py
Co-authored-by: shamoon <4887959+shamoon@users.noreply.github.com>
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.
Trenton H
shamoon
dependabot[bot]andTrenton Holmes