import json import logging import sqlite3 import struct from collections.abc import Iterator from collections.abc import Sequence from contextlib import contextmanager from pathlib import Path from types import TracebackType from typing import Any from typing import NamedTuple import sqlite_vec from llama_index.core.bridge.pydantic import PrivateAttr from llama_index.core.schema import BaseNode from llama_index.core.vector_stores.types import BasePydanticVectorStore from llama_index.core.vector_stores.types import FilterCondition from llama_index.core.vector_stores.types import FilterOperator from llama_index.core.vector_stores.types import MetadataFilter from llama_index.core.vector_stores.types import MetadataFilters from llama_index.core.vector_stores.types import VectorStoreQuery from llama_index.core.vector_stores.types import VectorStoreQueryResult from llama_index.core.vector_stores.utils import metadata_dict_to_node from llama_index.core.vector_stores.utils import node_to_metadata_dict from paperless_ai.migrations import MIGRATIONS from paperless_ai.migrations import Migration from paperless_ai.tables import ChunkRow from paperless_ai.tables import DocumentChunksTable from paperless_ai.tables import DocumentMetaRow from paperless_ai.tables import DocumentMetaTable from paperless_ai.tables import IndexMetaTable logger = logging.getLogger("paperless_ai.vector_store") DB_FILENAME = "llmindex.db" DEFAULT_TABLE_NAME = "documents" # Current schema version. Written to index_meta at table creation and bumped # whenever a Migration is added to MIGRATIONS. check_and_run_migrations() uses # this to decide which migrations to run on an existing store. SCHEMA_VERSION = 2 # compact(): rebuild when the cumulative rowid count exceeds this multiple of # the live row count. DELETEs on vec0 tables never reclaim space (upstream # asg017/sqlite-vec#54), so per-document re-index churn grows the file until # a rebuild copies the live rows into a fresh table. COMPACT_BLOAT_RATIO = 2.0 # compact(): number of rows copied per executemany() when rebuilding the file. # Rows are streamed from the source cursor in batches of this size rather than # materialized all at once, keeping memory bounded regardless of index size. COMPACT_BATCH_SIZE = 500 # Filterable vec0 metadata columns. _build_where() only ever receives filter # keys we construct ourselves, but allowlisting keeps SQL identifiers safe by # construction. "modified" is not here: it is never filtered on, and as of # schema v2 it isn't even a vec0 column anymore (see document_meta). _FILTER_COLUMNS = frozenset({"document_id"}) class _Row(NamedTuple): """One node, ready to write. ``modified`` is not a vec0 column (see document_meta) -- it rides along here because every row-producing call site needs both the vec0 insert values and the document_meta upsert value from the same node. """ chunk_id: str document_id: int modified: str node_content: str embedding: bytes def _pack(embedding: Sequence[float]) -> bytes: return struct.pack(f"{len(embedding)}f", *embedding) def _unpack(blob: bytes) -> list[float]: return list(struct.unpack(f"{len(blob) // 4}f", blob)) _INSERT = ( "INSERT INTO " + DEFAULT_TABLE_NAME + " (id, document_id, node_content, embedding) VALUES (?, ?, ?, ?)" ) def _vec0_params(rows: list[_Row]) -> list[tuple[str, int, str, bytes]]: """``rows``, minus the ``modified`` field vec0 no longer stores.""" return [(r.chunk_id, r.document_id, r.node_content, r.embedding) for r in rows] def _build_where(filters: MetadataFilters | None) -> tuple[str, list[int]]: """Translate the EQ / IN / NE filters we use into a parameterized SQL clause on vec0 metadata columns. Returns ("", []) when there is nothing to filter. document_id is vec0's only filterable column and is INTEGER; every value is coerced via int() here so callers (which today still pass strings in places, e.g. indexing.py's MetadataFilter construction) don't have to be individually correct -- vec0 doesn't coerce types itself. """ if filters is None or not filters.filters: return "", [] clauses: list[str] = [] params: list[int] = [] for f in filters.filters: # filters.filters is Union[MetadataFilter, ExactMatchFilter, MetadataFilters]; # we only build MetadataFilter entries, so skip anything else at runtime. if not isinstance(f, MetadataFilter): continue if f.key not in _FILTER_COLUMNS: # pragma: no cover - we build the keys raise NotImplementedError(f"Unsupported filter column: {f.key}") if f.operator == FilterOperator.IN: values = [int(v) for v in f.value] # type: ignore[union-attr] if not values: # pragma: no cover clauses.append("1 = 0") continue placeholders = ",".join("?" for _ in values) clauses.append(f"{f.key} IN ({placeholders})") params.extend(values) elif f.operator == FilterOperator.EQ: clauses.append(f"{f.key} = ?") params.append(int(f.value)) elif f.operator == FilterOperator.NE: clauses.append(f"{f.key} != ?") params.append(int(f.value)) else: # pragma: no cover - we only ever build EQ/IN/NE filters raise NotImplementedError(f"Unsupported filter operator: {f.operator}") if not clauses: # Filters were requested but none could be translated. Fail closed # rather than emit "()" (invalid SQL): filters scope document access, # so an empty translation must match no rows, never widen the scope. return "1 = 0", [] joiner = " OR " if filters.condition == FilterCondition.OR else " AND " return "(" + joiner.join(clauses) + ")", params class PaperlessSqliteVecVectorStore(BasePydanticVectorStore): """A llama-index vector store backed by a sqlite-vec vec0 table. Stores one row per node: the node id (TEXT primary key), its document id (metadata column, used for EQ/IN filtering and per-document delete), the document's modified timestamp, the embedding (float32, cosine metric), and the serialized node (text + metadata) as JSON in an auxiliary column. ``stores_text`` lets llama-index run off this store alone, with no separate docstore or index store. Everything lives in one SQLite database file (``DB_FILENAME``) inside the directory given as ``uri`` (kept as a directory for compatibility with the previous LanceDB layout). WAL mode allows readers in other processes to proceed while the (FileLock-serialized) writer holds a transaction. Implemented surface of ``BasePydanticVectorStore`` --------------------------------------------------- Only the methods actively used by this codebase are implemented. ``delete_nodes`` and the ``node_ids`` lookup path of ``get_nodes`` are part of the llama-index interface contract and may be needed if a future retriever or extension invokes them — add them then, with tests. """ stores_text: bool = True flat_metadata: bool = False _uri: str = PrivateAttr() _embed_model_name: str | None = PrivateAttr() _conn: Any = PrivateAttr() def __init__( self, uri: str, embed_model_name: str | None = None, ) -> None: super().__init__(stores_text=True, flat_metadata=False) self._uri = uri self._embed_model_name = embed_model_name self._conn = self._open_connection(str(Path(uri) / DB_FILENAME)) @staticmethod def _open_connection(db_path: str) -> sqlite3.Connection: conn = sqlite3.connect( db_path, timeout=30, isolation_level=None, # autocommit; explicit transactions below ) conn.row_factory = sqlite3.Row conn.enable_load_extension(True) # noqa: FBT003 sqlite_vec.load(conn) conn.enable_load_extension(False) # noqa: FBT003 conn.execute("PRAGMA journal_mode=WAL") conn.execute("PRAGMA synchronous=NORMAL") IndexMetaTable.create(conn) # vec0 metadata columns only get an efficient lookup path inside a # KNN (MATCH) query; a plain `WHERE document_id = ?` is a full table # scan regardless of index size. This plain, indexed table is how # delete()/upsert_document() find a document's chunk ids without # that scan. DocumentChunksTable.create(conn) # modified used to be a vec0 metadata column, but vec0 only inlines # TEXT metadata up to 12 bytes -- an ISO timestamp is always longer, # so every read recompiled and stepped a fresh SQL statement per row. # It was never filtered on inside a KNN query either, so it never # needed to be a vec0 column at all. One row per document here (not # per chunk, like document_chunks), since every chunk of a document # shares the same modified value -- see get_modified_times(). DocumentMetaTable.create(conn) return conn @property def client(self) -> Any: return self._conn def close(self) -> None: """Close the underlying SQLite connection (idempotent).""" self._conn.close() def __enter__(self) -> "PaperlessSqliteVecVectorStore": return self def __exit__( self, exc_type: type[BaseException] | None, exc_val: BaseException | None, exc_tb: TracebackType | None, ) -> None: # Deterministically release the connection (and its WAL/SHM handles) so # it is never left open across a compaction/migration file swap. self.close() @contextmanager def _transaction(self) -> Iterator[None]: self._conn.execute("BEGIN IMMEDIATE") try: yield except BaseException: # pragma: no cover self._conn.execute("ROLLBACK") raise else: self._conn.execute("COMMIT") def table_exists(self) -> bool: return ( self._conn.execute( "SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = ?", (DEFAULT_TABLE_NAME,), ).fetchone() is not None ) def vector_dim(self) -> int | None: if not self.table_exists(): return None return IndexMetaTable.get_dim(self._conn) def drop_table(self) -> None: self._conn.execute("DROP TABLE IF EXISTS " + DEFAULT_TABLE_NAME) self._conn.execute("DELETE FROM index_meta") DocumentChunksTable.delete_all(self._conn) DocumentMetaTable.delete_all(self._conn) def stored_model_name(self) -> str | None: """Return the embedding model name recorded at table creation, or None.""" if not self.table_exists(): return None return IndexMetaTable.get_embed_model(self._conn) def config_mismatch(self, model_name: str) -> bool: """True when the stored model name differs from ``model_name``. Returns False when no table exists or when the table predates model-name tracking — conservative default avoids spurious rebuilds. """ stored = self.stored_model_name() if stored is None: return False return stored != model_name @staticmethod def _create_vec_table(conn: sqlite3.Connection, dim: int) -> None: # document_id is deliberately a metadata column, NOT a partition key: # partition keys change KNN `k` to per-partition semantics under IN # filters (asg017/sqlite-vec#142); metadata columns give a correct # global top-k. INTEGER (not TEXT, as in schema v1): EQ/NE/IN # comparisons become a native i64 array compare instead of per-row # strncmp against a 16-byte text view, and this drops the unused # metadatatext shadow table TEXT columns carry. modified is not a # column here at all as of v2 -- see document_meta. conn.execute( # nosemgrep: python.sqlalchemy.security.sqlalchemy-execute-raw-query.sqlalchemy-execute-raw-query "CREATE VIRTUAL TABLE " + DEFAULT_TABLE_NAME + " USING vec0(" + "id TEXT PRIMARY KEY," + " document_id INTEGER," + " +node_content TEXT," + " embedding float[" + str(int(dim)) + "] distance_metric=cosine" + ")", ) def _create_table(self, dim: int) -> None: self._create_vec_table(self._conn, dim) IndexMetaTable.set_dim(self._conn, dim) IndexMetaTable.set_schema_version(self._conn, SCHEMA_VERSION) if self._embed_model_name: IndexMetaTable.set_embed_model(self._conn, self._embed_model_name) def _ensure_table(self, dim: int) -> None: if not self.table_exists(): self._create_table(dim) def _row(self, node: BaseNode) -> _Row: meta = node_to_metadata_dict( node, remove_text=False, flat_metadata=self.flat_metadata, ) document_id = node.ref_doc_id or node.metadata.get("document_id") return _Row( chunk_id=node.node_id, document_id=int(document_id), modified=str(node.metadata.get("modified") or ""), node_content=json.dumps(meta), embedding=_pack(node.get_embedding()), ) def _index_chunks(self, rows: list[_Row]) -> None: """Record each row's (chunk_id, document_id) in document_chunks, and each row's (document_id, modified) in document_meta -- deduped within the batch, since every chunk of a document shares the same modified value -- kept in lockstep with every insert into the vec0 table. """ DocumentChunksTable.insert_many( self._conn, (ChunkRow(r.chunk_id, r.document_id) for r in rows), ) modified_by_document = {r.document_id: r.modified for r in rows} DocumentMetaTable.upsert_many( self._conn, ( DocumentMetaRow(doc_id, mod) for doc_id, mod in modified_by_document.items() ), ) def _delete_chunks_by_document_id(self, document_id: int) -> None: """Delete all of a document's chunks via point-deletes on `id`. vec0 has no efficient lookup on the document_id metadata column outside a KNN query, so a plain `DELETE ... WHERE document_id = ?` is a full table scan regardless of index size. Looking the chunk ids up in document_chunks first (a real indexed lookup) and deleting each by its `id` primary key instead turns that scan into a handful of O(1) point deletes. """ chunk_ids = DocumentChunksTable.chunk_ids_for_document( self._conn, document_id, ) self._conn.executemany( "DELETE FROM " + DEFAULT_TABLE_NAME + " WHERE id = ?", [(chunk_id,) for chunk_id in chunk_ids], ) DocumentChunksTable.delete_for_document(self._conn, document_id) DocumentMetaTable.delete_for_document(self._conn, document_id) def _increment_total_inserts(self, count: int) -> None: """Increment the cumulative insert counter stored in index_meta. This counter never decreases (DELETEs do not decrement it) and is used by compact() to estimate the bloat ratio: when total_inserts / live_rows exceeds COMPACT_BLOAT_RATIO the table has accumulated enough deleted-but-not-freed rows to warrant a rebuild. """ IndexMetaTable.increment_total_inserts(self._conn, count) def add(self, nodes: Sequence[BaseNode], **add_kwargs: Any) -> list[str]: if not nodes: return [] rows = [self._row(node) for node in nodes] with self._transaction(): self._ensure_table(len(nodes[0].get_embedding())) self._conn.executemany(_INSERT, _vec0_params(rows)) self._index_chunks(rows) self._increment_total_inserts(len(rows)) return [node.node_id for node in nodes] def upsert_document( self, document_id: int | str, nodes: list[BaseNode], ) -> list[str]: """Atomically replace all stored chunks of ``document_id`` with ``nodes``. One transaction deletes the document's existing rows and inserts the new set (vec0's INSERT OR REPLACE is broken upstream, so delete+insert it is). WAL readers in other processes see either the old or the new chunk set, never a partial state. """ doc_id = int(document_id) rows = [self._row(node) for node in nodes] with self._transaction(): if nodes: self._ensure_table(len(nodes[0].get_embedding())) if self.table_exists(): self._delete_chunks_by_document_id(doc_id) if rows: self._conn.executemany(_INSERT, _vec0_params(rows)) self._index_chunks(rows) self._increment_total_inserts(len(rows)) return [node.node_id for node in nodes] def delete(self, ref_doc_id: int | str, **delete_kwargs: Any) -> None: if self.table_exists(): with self._transaction(): self._delete_chunks_by_document_id(int(ref_doc_id)) def _rows_to_nodes(self, rows: list[sqlite3.Row]) -> list[BaseNode]: nodes: list[BaseNode] = [] for row in rows: node = metadata_dict_to_node(json.loads(row["node_content"])) node.embedding = _unpack(row["embedding"]) nodes.append(node) return nodes def get_nodes( self, node_ids: list[str] | None = None, filters: MetadataFilters | None = None, **kwargs: Any, ) -> list[BaseNode]: if node_ids is not None: # pragma: no cover # node_ids lookup is not implemented; see class docstring. raise NotImplementedError( "PaperlessSqliteVecVectorStore does not support node_ids lookup", ) if not self.table_exists(): return [] where, params = _build_where(filters) sql = "SELECT node_content, embedding FROM " + DEFAULT_TABLE_NAME if where: sql += " WHERE " + where return self._rows_to_nodes(self._conn.execute(sql, params).fetchall()) def query( self, query: VectorStoreQuery, **kwargs: Any, ) -> VectorStoreQueryResult: if not self.table_exists(): return VectorStoreQueryResult(nodes=[], similarities=[], ids=[]) if query.query_embedding is None: # pragma: no cover return VectorStoreQueryResult(nodes=[], similarities=[], ids=[]) top_k = query.similarity_top_k if query.similarity_top_k is not None else 10 where, params = _build_where(query.filters) sql = ( "SELECT id, node_content, embedding, distance FROM " + DEFAULT_TABLE_NAME + " WHERE embedding MATCH ? AND k = ?" ) if where: sql += " AND " + where rows = self._conn.execute( sql, [_pack(query.query_embedding), top_k, *params], ).fetchall() # vec0 returns rows distance-sorted ascending; slice defensively in # case future schema changes alter k semantics (e.g. partition keys # return k rows per partition). rows = rows[:top_k] nodes = self._rows_to_nodes(rows) # Cosine distance in [0, 2]; map to a descending similarity. # vec0 returns None distance when the query embedding is the zero vector # (no meaningful cosine angle); treat that as maximum distance (1.0) so # the row is included but ranked last. sims = [ 1.0 - float(row["distance"] if row["distance"] is not None else 1.0) for row in rows ] ids = [row["id"] for row in rows] return VectorStoreQueryResult(nodes=nodes, similarities=sims, ids=ids) def get_modified_times(self) -> dict[str, str]: """Return {document_id: stored_modified_isoformat} for all indexed documents. document_meta already has exactly one row per document (not per chunk, unlike the vec0 table), so no dedup is needed here. """ if not self.table_exists(): return {} return DocumentMetaTable.all_modified_times(self._conn) @property def _db_path(self) -> str: return str(Path(self._uri) / DB_FILENAME) @contextmanager def _rebuild_file(self) -> Iterator[sqlite3.Connection]: """Open a fresh temp database file for a file-swap rebuild (compact or structural migration), yielding its connection for the caller to populate. On success, swaps the temp file in as the live database (closing this store's current connection first -- see _swap_in_compact()). On any exception, discards the temp file, including its -wal/-shm, instead, and this store's own connection is left untouched. """ compact_path = self._db_path + ".compact" new_conn = self._open_connection(compact_path) try: yield new_conn except BaseException: new_conn.close() for suffix in ["", "-wal", "-shm"]: Path(compact_path + suffix).unlink(missing_ok=True) raise else: new_conn.close() self._swap_in_compact(compact_path, self._db_path) def compact(self, *, force: bool = False) -> None: """Rebuild the database file to reclaim space left behind by DELETEs. vec0 DELETE only invalidates rows; the vector data stays in the file forever, and per-document re-indexing is a delete+insert. The cumulative insert counter in ``index_meta`` tracks total rows ever written; when that exceeds ``COMPACT_BLOAT_RATIO`` x the live row count (or when forced), live rows are copied into a fresh database file and swapped in via ``os.replace``. Note: ``ALTER TABLE ... RENAME TO`` on vec0 virtual tables does NOT rename the shadow tables (sqlite-vec upstream limitation), so an in-place rename-based rebuild is not safe. The file-swap approach is the maintainer-endorsed workaround. """ if not self.table_exists(): return live = DocumentChunksTable.count(self._conn) total = IndexMetaTable.get_total_inserts(self._conn) or live if not force and total <= max(live, 1) * COMPACT_BLOAT_RATIO: return dim = self.vector_dim() if dim is None: # pragma: no cover - dim is written at creation logger.warning("Skipping compact: no stored vector dimension") return logger.info( "Compacting LLM index (%d live rows, %d cumulative inserts)", live, total, ) with self._rebuild_file() as new_conn: self._rebuild_into(self._conn, new_conn, dim) @staticmethod def _rebuild_into( src_conn: sqlite3.Connection, dst_conn: sqlite3.Connection, dim: int, ) -> int: """Create the vec0 table in ``dst_conn``, copy dim/embed_model from ``src_conn``, and stream every live vec0 row, document_chunks row, and document_meta row across. Returns the number of vec0 rows copied. Used by compact() only -- m0001_v1_to_v2 freezes its own copy loop instead of calling this, since this always reflects the *current* schema (see the migration DDL-freezing rule in the spec). """ PaperlessSqliteVecVectorStore._create_vec_table(dst_conn, dim) dim_value = IndexMetaTable.get_dim(src_conn) if dim_value is not None: IndexMetaTable.set_dim(dst_conn, dim_value) embed_model = IndexMetaTable.get_embed_model(src_conn) if embed_model is not None: IndexMetaTable.set_embed_model(dst_conn, embed_model) schema_version = IndexMetaTable.get_schema_version(src_conn) if schema_version is not None: IndexMetaTable.set_schema_version(dst_conn, schema_version) dst_conn.execute("BEGIN IMMEDIATE") src_cursor = src_conn.execute( "SELECT id, document_id, node_content, embedding FROM " + DEFAULT_TABLE_NAME, ) copied = 0 while batch := src_cursor.fetchmany(COMPACT_BATCH_SIZE): dst_conn.executemany( _INSERT, [ ( r["id"], r["document_id"], r["node_content"], bytes(r["embedding"]), ) for r in batch ], ) DocumentChunksTable.insert_many( dst_conn, (ChunkRow(r["id"], r["document_id"]) for r in batch), ) copied += len(batch) DocumentMetaTable.copy_all(src_conn, dst_conn, COMPACT_BATCH_SIZE) # Reset the cumulative counter: after a rebuild, total_inserts == live. IndexMetaTable.reset_total_inserts(dst_conn, copied) dst_conn.execute("COMMIT") return copied def _swap_in_compact(self, compact_path: str, db_path: str) -> None: """Atomically replace the live database with the compacted copy.""" self._conn.close() for suffix in ["-wal", "-shm"]: stale = Path(compact_path + suffix) if stale.exists(): # pragma: no cover stale.unlink() Path(compact_path).replace(db_path) self._conn = self._open_connection(db_path) def _stored_schema_version(self) -> int | None: """The schema_version recorded in index_meta, or None if no table exists. A missing key (a store predating version tracking) is treated as SCHEMA_VERSION -- i.e. already current -- since no migration in MIGRATIONS targets a version before tracking began. """ if not self.table_exists(): return None raw_version = IndexMetaTable.get_schema_version(self._conn) return raw_version if raw_version is not None else SCHEMA_VERSION def has_pending_migration(self) -> bool: """Cheaply check whether a migration is pending, with no exclusive access needed -- just a metadata read under the connection callers already hold via the write FileLock. Callers should only pay for check_and_run_migrations()'s exclusive access (a structural migration's file swap must not run while readers are active) when this returns True, so that the common case -- already at SCHEMA_VERSION -- never contends with readers or a concurrent compaction. """ current = self._stored_schema_version() return current is not None and current < SCHEMA_VERSION def check_and_run_migrations(self) -> bool: """Apply any pending schema migrations to the store. Structural migrations copy live rows into a new-schema file with no re-embedding. Re-embed migrations cannot be applied automatically; this method returns True when one is encountered so the caller can force a full rebuild (which recreates the table at SCHEMA_VERSION). Must be called under the write FileLock, with readers excluded (see has_pending_migration() for a cheap pre-check that avoids paying for that exclusion in the common case). No-op when the table does not exist or is already at SCHEMA_VERSION. """ current = self._stored_schema_version() if current is None or current >= SCHEMA_VERSION: return False pending = sorted( [m for m in MIGRATIONS if current <= m.from_version < SCHEMA_VERSION], key=lambda m: m.from_version, ) for migration in pending: if migration.kind == "re-embed": logger.warning( "LLM index schema v%d -> v%d requires re-embedding (%s); " "the caller must force a rebuild.", migration.from_version, migration.to_version, migration.description, ) return True logger.info( "Running structural LLM index migration v%d -> v%d: %s", migration.from_version, migration.to_version, migration.description, ) self._run_structural_migration(migration) return False def _run_structural_migration(self, migration: Migration) -> None: """Execute a structural migration using the same file-swap as compact().""" assert migration.apply is not None, "structural migration must have apply()" dim = self.vector_dim() if dim is None: # pragma: no cover raise RuntimeError("Cannot migrate: no stored vector dimension") with self._rebuild_file() as new_conn: migration.apply(self._conn, new_conn, dim) IndexMetaTable.set_schema_version(new_conn, migration.to_version) # Registers m0001_v1_to_v2 into MIGRATIONS; must be at the bottom (needs # PaperlessSqliteVecVectorStore fully defined) -- see # paperless_ai/migrations/__init__.py for the full procedure. from paperless_ai.migrations import m0001_v1_to_v2 # noqa: E402, F401