basic-memory 0.23.0__tar.gz → 0.23.1.dev20__tar.gz

basic_memory-0.23.1.dev20/.github/workflows/mcp-registry-publish.yml

@@ -0,0 +1,36 @@
+name: MCP Registry Publish
+
+# Publishes server.json to the official MCP Registry. GitHub Actions OIDC
+# authenticates this workflow as the repository owner, which grants the
+# io.github.basicmachines-co/* namespace directly — no personal login or
+# public-org-membership derivation involved (the manual `mcp-publisher login
+# github` path proved flaky there; see the v0.23.0 release notes on #1288).
+#
+# Run manually after a release, once the version is live on PyPI.
+
+on:
+  workflow_dispatch:
+
+env:
+  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      id-token: write
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Install mcp-publisher
+        run: |
+          curl -fsSL "https://github.com/modelcontextprotocol/registry/releases/latest/download/mcp-publisher_$(uname -s | tr '[:upper:]' '[:lower:]')_$(uname -m | sed 's/x86_64/amd64/').tar.gz" \
+            | tar xz mcp-publisher
+          sudo mv mcp-publisher /usr/local/bin/
+
+      - name: Publish server.json
+        run: |
+          mcp-publisher login github-oidc
+          mcp-publisher publish

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: basic-memory
-Version: 0.23.0
+Version: 0.23.1.dev20
 Summary: Local-first knowledge management combining Zettelkasten with knowledge graphs
 Project-URL: Homepage, https://github.com/basicmachines-co/basic-memory
 Project-URL: Repository, https://github.com/basicmachines-co/basic-memory

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/docs/NOTE-FORMAT.md

@@ -166,6 +166,17 @@ Explicit relations:
 - 'in response to' [[Incident Review]]
 ```
 
+When the text before `[[` is ordinary prose rather than a relation type, add
+`#bm:links_to` to force the reference to use the implicit `links_to` relation:
+
+```markdown
+- Mother [[Alice]] #bm:links_to
+```
+
+The directive is parser metadata and is not included in the observation or
+relation context. It is especially useful when a single-token prose prefix
+would otherwise be interpreted as an explicit relation type.
+
 Bare wiki links and prose list items create implicit `links_to` relations:
 
 ```markdown

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/justfile

@@ -521,7 +521,6 @@ doctor:
     TMP_HOME=$(mktemp -d)
     TMP_CONFIG=$(mktemp -d)
     HOME="$TMP_HOME" \
-    BASIC_MEMORY_ENV=test \
     BASIC_MEMORY_HOME="$TMP_HOME/basic-memory" \
     BASIC_MEMORY_CONFIG_DIR="$TMP_CONFIG" \
     ./.venv/bin/python -m basic_memory.cli.main doctor --local
@@ -706,6 +705,8 @@ release version:
         .claude-plugin/marketplace.json \
         plugins/claude-code/.claude-plugin/plugin.json \
         plugins/claude-code/.claude-plugin/marketplace.json \
+        plugins/claude-code/hooks/session_start.py \
+        plugins/claude-code/hooks/pre_compact.py \
         plugins/codex/.codex-plugin/plugin.json \
         plugins/codex/hooks/session_start.py \
         plugins/codex/hooks/pre_compact.py \
@@ -826,6 +827,8 @@ beta version:
         .claude-plugin/marketplace.json \
         plugins/claude-code/.claude-plugin/plugin.json \
         plugins/claude-code/.claude-plugin/marketplace.json \
+        plugins/claude-code/hooks/session_start.py \
+        plugins/claude-code/hooks/pre_compact.py \
         plugins/codex/.codex-plugin/plugin.json \
         plugins/codex/hooks/session_start.py \
         plugins/codex/hooks/pre_compact.py \

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/plugins/claude-code/hooks/pre_compact.py

@@ -1,7 +1,14 @@
 #!/usr/bin/env -S uv run --quiet --script
 # /// script
 # requires-python = ">=3.12"
-# dependencies = ["basic-memory>=0.22.1"]
+# dependencies = [
+#     "basic-memory>=0.23.0",
+#     # Direct pre-release pin so old uv resolves basic-memory (which
+#     # requires this exact beta transitively) without enabling
+#     # pre-releases for basic-memory itself. Keep in lockstep with
+#     # core pyproject's fastmcp pin.
+#     "fastmcp==4.0.0b1",
+# ]
 # ///
 """PreCompact hook — the entire hook. All logic (settings resolution, the
 extractive checkpoint note, lifecycle-envelope capture) lives in the released

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/plugins/claude-code/hooks/session_start.py

@@ -1,7 +1,14 @@
 #!/usr/bin/env -S uv run --quiet --script
 # /// script
 # requires-python = ">=3.12"
-# dependencies = ["basic-memory>=0.22.1"]
+# dependencies = [
+#     "basic-memory>=0.23.0",
+#     # Direct pre-release pin so old uv resolves basic-memory (which
+#     # requires this exact beta transitively) without enabling
+#     # pre-releases for basic-memory itself. Keep in lockstep with
+#     # core pyproject's fastmcp pin.
+#     "fastmcp==4.0.0b1",
+# ]
 # ///
 """SessionStart hook — the entire hook. All logic (settings resolution, the
 context brief, lifecycle-envelope capture) lives in the released basic-memory

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/plugins/claude-code/hooks/test_claude_pre_compact.py

@@ -119,7 +119,7 @@ def test_dependency_floor_matches_package_version() -> None:
     # scripts/update_versions.py bumps this line at release; drift between the
     # script floor and the package version fails here, before a release lands.
     text = SCRIPT.read_text(encoding="utf-8")
-    floors = re.findall(r'^# dependencies = \["basic-memory>=([^"]+)"\]$', text, re.MULTILINE)
+    floors = re.findall(r'^#     "basic-memory>=([^"]+)",$', text, re.MULTILINE)
 
     assert floors == [__version__]
 

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/plugins/claude-code/hooks/test_claude_session_start.py

@@ -119,7 +119,7 @@ def test_dependency_floor_matches_package_version() -> None:
     # scripts/update_versions.py bumps this line at release; drift between the
     # script floor and the package version fails here, before a release lands.
     text = SCRIPT.read_text(encoding="utf-8")
-    floors = re.findall(r'^# dependencies = \["basic-memory>=([^"]+)"\]$', text, re.MULTILINE)
+    floors = re.findall(r'^#     "basic-memory>=([^"]+)",$', text, re.MULTILINE)
 
     assert floors == [__version__]
 

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/plugins/codex/hooks/pre_compact.py

@@ -3,6 +3,10 @@
 # requires-python = ">=3.12"
 # dependencies = [
 #     "basic-memory @ git+https://github.com/basicmachines-co/basic-memory@v0.23.0",
+#     # Direct pre-release pin so old uv resolves basic-memory (which
+#     # requires this exact beta transitively) without enabling
+#     # pre-releases broadly. Keep in lockstep with core pyproject.
+#     "fastmcp==4.0.0b1",
 # ]
 # ///
 """PreCompact hook launcher backed by a pinned Basic Memory revision.

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/plugins/codex/hooks/session_start.py

@@ -3,6 +3,10 @@
 # requires-python = ">=3.12"
 # dependencies = [
 #     "basic-memory @ git+https://github.com/basicmachines-co/basic-memory@v0.23.0",
+#     # Direct pre-release pin so old uv resolves basic-memory (which
+#     # requires this exact beta transitively) without enabling
+#     # pre-releases broadly. Keep in lockstep with core pyproject.
+#     "fastmcp==4.0.0b1",
 # ]
 # ///
 """SessionStart hook launcher backed by a pinned Basic Memory revision.

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/scripts/update_versions.py

@@ -153,8 +153,8 @@ def _update_packages(version: str, *, dry_run: bool) -> None:
     for script in HOOK_SCRIPTS:
         update_text(
             script,
-            r'^# dependencies = \["basic-memory>=[^"]+"\]$',
-            f'# dependencies = ["basic-memory>={version}"]',
+            r'^#     "basic-memory>=[^"]+",$',
+            f'#     "basic-memory>={version}",',
             dry_run=dry_run,
         )
     update_text(

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/scripts/validate_claude_plugin.py

@@ -137,6 +137,20 @@ def validate_claude_plugin(plugin_dir: Path) -> None:
     for event in REQUIRED_HOOK_EVENTS:
         if event not in hooks:
             raise SystemExit(f"hooks/hooks.json: missing {event} hook")
+    # Trigger: the shims pin core's exact fastmcp beta so older uv can resolve
+    #   basic-memory (pre-release transitives are refused there) without
+    #   enabling pre-releases for basic-memory itself.
+    # Why: nothing else keeps the shim pin and core's pyproject in lockstep,
+    #   and drift makes shim resolution conflict — which the fail-open contract
+    #   turns into a silent hook no-op.
+    # Outcome: validation fails loudly when core moves its fastmcp pin without
+    #   the shims following.
+    core_pyproject = (ROOT / "pyproject.toml").read_text(encoding="utf-8")
+    core_match = re.search(r'^\s*"fastmcp==([^"]+)",$', core_pyproject, re.MULTILINE)
+    if not core_match:
+        raise SystemExit("pyproject.toml: no exact fastmcp pin found")
+    core_fastmcp_pin = core_match.group(1)
+
     for rel in REQUIRED_HOOK_SCRIPTS:
         script = plugin_dir / rel
         if not script.exists():
@@ -145,9 +159,17 @@ def validate_claude_plugin(plugin_dir: Path) -> None:
             raise SystemExit(f"Hook script is not executable: {script}")
         text = script.read_text(encoding="utf-8")
         if "# /// script" not in text or not re.search(
-            r'^# dependencies = \["basic-memory>=[^"]+"\]$', text, re.MULTILINE
+            r'^#     "basic-memory>=[^"]+",$', text, re.MULTILINE
         ):
             raise SystemExit(f"Hook script missing PEP 723 basic-memory floor: {script}")
+        shim_pins = re.findall(r'^#     "fastmcp==([^"]+)",$', text, re.MULTILINE)
+        if len(shim_pins) != 1:
+            raise SystemExit(f"Hook script missing exact fastmcp pin: {script}")
+        if shim_pins[0] != core_fastmcp_pin:
+            raise SystemExit(
+                f"{script}: fastmcp pin {shim_pins[0]} does not match "
+                f"core pyproject pin {core_fastmcp_pin}"
+            )
 
     # --- Output style ---
     output_style = plugin_dir / "output-styles/basic-memory.md"

basic_memory-0.23.1.dev20/src/basic_memory/alembic/versions/7f6a2b8c9d10_index_full_postgres_note_content.py

@@ -0,0 +1,77 @@
+"""Index complete PostgreSQL note content for full-text search.
+
+Revision ID: 7f6a2b8c9d10
+Revises: 2d26b287813b
+Create Date: 2026-08-24 22:00:00.000000
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+# revision identifiers, used by Alembic.
+revision: str = "7f6a2b8c9d10"
+down_revision: Union[str, None] = "2d26b287813b"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Index complete note bodies as bounded PostgreSQL FTS chunks."""
+    connection = op.get_bind()
+    if connection.dialect.name == "postgresql":
+        op.execute("""
+            CREATE TABLE search_index_fts_chunks (
+                project_id INTEGER NOT NULL,
+                search_index_id INTEGER NOT NULL,
+                search_index_type VARCHAR NOT NULL,
+                chunk_index INTEGER NOT NULL,
+                chunk_text TEXT NOT NULL,
+                textsearchable_index_col tsvector GENERATED ALWAYS AS (
+                    to_tsvector('english', chunk_text)
+                ) STORED,
+                PRIMARY KEY (project_id, search_index_id, search_index_type, chunk_index),
+                FOREIGN KEY (search_index_id, search_index_type, project_id)
+                    REFERENCES search_index(id, type, project_id)
+                    ON UPDATE CASCADE ON DELETE CASCADE
+            )
+        """)
+        op.execute("""
+            -- PostgreSQL ignores lexemes at 2 KiB and above. Stepping 5,952
+            -- characters leaves a conservative 2,048-character overlap, so
+            -- every indexable lexeme split at one edge is complete in the next.
+            INSERT INTO search_index_fts_chunks (
+                project_id,
+                search_index_id,
+                search_index_type,
+                chunk_index,
+                chunk_text
+            )
+            SELECT
+                search_index.project_id,
+                search_index.id,
+                search_index.type,
+                (chunk_start - 1) / 5952,
+                substring(search_index.content_snippet FROM chunk_start FOR 8000)
+            FROM search_index
+            CROSS JOIN LATERAL generate_series(
+                1,
+                length(search_index.content_snippet),
+                5952
+            ) AS chunk_start
+            WHERE search_index.content_snippet IS NOT NULL
+              AND search_index.content_snippet <> ''
+        """)
+        op.execute("""
+            CREATE INDEX idx_search_index_fts_chunks_fts
+            ON search_index_fts_chunks USING gin(textsearchable_index_col)
+        """)
+
+
+def downgrade() -> None:
+    """Remove the bounded full-content PostgreSQL FTS chunks."""
+    connection = op.get_bind()
+    if connection.dialect.name == "postgresql":
+        op.execute("DROP TABLE IF EXISTS search_index_fts_chunks")

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/src/basic_memory/cli/auto_update.py

@@ -221,6 +221,38 @@ def _preload_lazy_console_modules() -> None:
     """
     import rich._emoji_codes  # noqa: F401
     import typer.rich_utils  # noqa: F401
+    from rich.cells import cell_len
+
+    # Trigger: rich defers its Unicode cell-width table (`rich._unicode_data.
+    # unicode<version>`) until the first character it cannot measure with the
+    # ASCII fast path in `_cell_len`.
+    # Why: status messages echo captured `brew`/`uv` output, which carries
+    # non-ASCII characters (curly quotes, em dashes, warning glyphs), so the
+    # deferred import lands after the upgrade removed our files. Importing the
+    # module by name would hard-code a table version; calling `cell_len` uses
+    # rich's own resolution and honors UNICODE_VERSION like the print path does.
+    # Outcome: the table rich will reach for is resolved and cached up front.
+    cell_len("\u2500\u2018\u2713")
+
+
+def print_update_status(console: Console, text: str, style: str) -> None:
+    """Print an update status line that cannot fail the command.
+
+    Trigger: the line is printed after an in-place upgrade may already have
+    replaced this installation on disk.
+    Why: `_preload_lazy_console_modules` can only preload the deferred imports
+    we know about today, and rich/typer are free to add more. By the time this
+    prints, the upgrade has already succeeded -- a status line must never be
+    what turns it into a traceback and a non-zero exit.
+    Outcome: fall back to a plain, unstyled write that needs no new imports.
+    """
+    try:
+        console.print(f"[{style}]{text}[/{style}]")
+    except Exception as exc:
+        logger.warning(
+            f"Rich console print failed after update, falling back to plain output: {exc}"
+        )
+        print(text)
 
 
 def _save_last_checked_timestamp(config_manager: ConfigManager, checked_at: datetime) -> None:
@@ -468,11 +500,11 @@ def maybe_run_periodic_auto_update(
     }:
         out = console or Console()
         if result.status == AutoUpdateStatus.UPDATED:
-            out.print(f"[green]{result.message}[/green]")
+            print_update_status(out, f"{result.message}", "green")
         elif result.status == AutoUpdateStatus.FAILED:
             error_detail = f" {result.error}" if result.error else ""
-            out.print(f"[yellow]{result.message}{error_detail}[/yellow]")
+            print_update_status(out, f"{result.message}{error_detail}", "yellow")
         elif result.message:
-            out.print(f"[cyan]{result.message}[/cyan]")
+            print_update_status(out, f"{result.message}", "cyan")
 
     return result

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/src/basic_memory/cli/commands/doctor.py

@@ -81,6 +81,23 @@ async def _delete_doctor_project(
         await _delete_doctor_project_locally(project_name, project_id)
 
 
+async def _read_materialized_api_note(api_file: Path, file_path: str) -> str:
+    """Wait for the accepted local write, then read its canonical file."""
+    # Deferred: importing the materialization runtime at CLI module load would slow every
+    # command, while doctor is the only command that needs to observe its write immediately.
+    from basic_memory.index.note_content_materialization import drain_pending_materializations
+
+    # Trigger: production accepts note content before its markdown file is materialized.
+    # Why: doctor verifies the complete DB -> file contract, not only write acceptance.
+    # Outcome: wait for the same deferred work drained at normal CLI shutdown before checking.
+    await drain_pending_materializations()
+
+    if not api_file.exists():
+        raise ValueError(f"API note file missing: {file_path}")
+
+    return api_file.read_text(encoding="utf-8")
+
+
 async def run_doctor() -> None:
     """Run local consistency checks for file <-> database flows."""
     # Deferred: the markdown parsing stack is only needed while the checks run,
@@ -132,10 +149,7 @@ async def run_doctor() -> None:
                 api_result = await knowledge_client.create_entity(api_note.model_dump())
 
                 api_file = project_path / api_result.file_path
-                if not api_file.exists():
-                    raise ValueError(f"API note file missing: {api_result.file_path}")
-
-                api_text = api_file.read_text(encoding="utf-8")
+                api_text = await _read_materialized_api_note(api_file, api_result.file_path)
                 if api_note_title not in api_text:
                     raise ValueError("API note content missing from file")
 

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/src/basic_memory/cli/commands/update.py

@@ -4,7 +4,7 @@ import typer
 from rich.console import Console
 
 from basic_memory.cli.app import app
-from basic_memory.cli.auto_update import AutoUpdateStatus, run_auto_update
+from basic_memory.cli.auto_update import AutoUpdateStatus, print_update_status, run_auto_update
 
 console = Console()
 
@@ -22,19 +22,21 @@ def update(
 
     if result.status == AutoUpdateStatus.FAILED:
         detail = f" {result.error}" if result.error else ""
-        console.print(f"[red]{result.message or 'Update failed.'}{detail}[/red]")
+        print_update_status(console, f"{result.message or 'Update failed.'}{detail}", "red")
         raise typer.Exit(1)
 
     if result.status == AutoUpdateStatus.UPDATED:
-        console.print(f"[green]{result.message or 'Basic Memory updated successfully.'}[/green]")
+        print_update_status(
+            console, f"{result.message or 'Basic Memory updated successfully.'}", "green"
+        )
         return
 
     if result.status == AutoUpdateStatus.UP_TO_DATE:
-        console.print(f"[green]{result.message or 'Basic Memory is up to date.'}[/green]")
+        print_update_status(console, f"{result.message or 'Basic Memory is up to date.'}", "green")
         return
 
     if result.status == AutoUpdateStatus.UPDATE_AVAILABLE:
-        console.print(f"[cyan]{result.message or 'Update available.'}[/cyan]")
+        print_update_status(console, f"{result.message or 'Update available.'}", "cyan")
         return
 
-    console.print(f"[dim]{result.message or 'No update action was performed.'}[/dim]")
+    print_update_status(console, f"{result.message or 'No update action was performed.'}", "dim")

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/src/basic_memory/markdown/plugins.py

@@ -13,6 +13,15 @@ from markdown_it.token import Token
 # those bracket prefixes stay ordinary content (issue #1219).
 _TIMESTAMP_VALUE = r"\d{1,3}:\d{2}(?::\d{2})?(?:[.,]\d{1,3})?"
 _TIMESTAMP_CATEGORY = re.compile(rf"^{_TIMESTAMP_VALUE}(?:\s+-\s+{_TIMESTAMP_VALUE})?$")
+_LINKS_TO_DIRECTIVE = re.compile(r"\s+#bm:links_to\s*$")
+
+
+def remove_links_to_directive(content: str) -> tuple[str, bool]:
+    """Remove an exact terminal ``#bm:links_to`` directive from content."""
+    match = _LINKS_TO_DIRECTIVE.search(content)
+    if not match:
+        return content, False
+    return content[: match.start()].rstrip(), True
 
 
 def _is_task_marker_category(category: str) -> bool:
@@ -47,6 +56,7 @@ def is_observation(token: Token) -> bool:
         return False
     # Use token.tag which contains the actual content for test tokens, fallback to content
     content = (token.tag or token.content).strip()
+    content, _ = remove_links_to_directive(content)
     if not content:  # pragma: no cover
         return False
     # if it's a markdown_task, return false
@@ -74,6 +84,7 @@ def parse_observation(token: Token) -> Dict[str, Any]:
 
     # Use token.tag which contains the actual content for test tokens, fallback to content
     content = (token.tag or token.content).strip()
+    content, _ = remove_links_to_directive(content)
 
     # Parse [category] with regex; a timestamp-shaped prefix is not a category, so a
     # hashtag-promoted transcript line keeps its timecode inside the content instead.
@@ -325,17 +336,19 @@ def relation_plugin(md: MarkdownIt) -> None:
 
             # Only process inline tokens
             if token.type == "inline":
+                content = token.tag or token.content
+                content_without_directive, has_directive = remove_links_to_directive(content)
+
                 # Check for explicit relations in list items
-                if in_list_item and is_explicit_relation(token):
+                if in_list_item and not has_directive and is_explicit_relation(token):
                     rel = parse_relation(token)
                     if rel:
                         token.meta["relations"] = [rel]
 
                 # Always check for inline links in any text
                 else:
-                    content = token.tag or token.content
                     if "[[" in content:
-                        rels = parse_inline_relations(content)
+                        rels = parse_inline_relations(content_without_directive)
                         if rels:
                             token.meta["relations"] = token.meta.get("relations", []) + rels
 

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/src/basic_memory/mcp/tools/write_note.py

@@ -109,12 +109,15 @@ async def write_note(
         - Explicit: `- relation_type [[Entity]] (optional context)`
         - Quoted: `- "multi word relation type" [[Entity]] (optional context)`
         - Quoted: `- 'multi word relation type' [[Entity]] (optional context)`
+        - Disambiguation: Add `#bm:links_to` when prose before `[[Entity]]`
+          must not be treated as a single-token relation type
         - Inline: Any other `[[Entity]]` reference creates a `links_to` relation
 
         Examples:
         `- depends_on [[Content Parser]] (Need for semantic extraction)`
         `- "based on" [[Design Notes]]`
         `- 'in response to' [[Incident Review]]`
+        `- Mother [[Alice]] #bm:links_to`
         `- implements [[Search Spec]] (Initial implementation)`
         `- This feature extends [[Base Design]] and uses [[Core Utils]]`
 

{basic_memory-0.23.0 → basic_memory-0.23.1.dev20}/src/basic_memory/models/search.py

@@ -33,7 +33,11 @@ CREATE TABLE IF NOT EXISTS search_index (
     created_at TIMESTAMP WITH TIME ZONE,
     updated_at TIMESTAMP WITH TIME ZONE,
     textsearchable_index_col tsvector GENERATED ALWAYS AS (
-        to_tsvector('english', coalesce(title, '') || ' ' || coalesce(content_stems, ''))
+        to_tsvector(
+            'english',
+            coalesce(title, '') || ' ' ||
+            coalesce(content_stems, '')
+        )
     ) STORED,
     PRIMARY KEY (id, type, project_id),
     FOREIGN KEY (project_id) REFERENCES project(id) ON DELETE CASCADE
@@ -44,6 +48,30 @@ CREATE_POSTGRES_SEARCH_INDEX_FTS = DDL("""
 CREATE INDEX IF NOT EXISTS idx_search_index_fts ON search_index USING gin(textsearchable_index_col)
 """)
 
+# Full note bodies are stored in bounded child rows so one unusually large note
+# cannot exceed PostgreSQL's per-tsvector size limit.
+CREATE_POSTGRES_SEARCH_INDEX_FTS_CHUNKS_TABLE = DDL("""
+CREATE TABLE IF NOT EXISTS search_index_fts_chunks (
+    project_id INTEGER NOT NULL,
+    search_index_id INTEGER NOT NULL,
+    search_index_type VARCHAR NOT NULL,
+    chunk_index INTEGER NOT NULL,
+    chunk_text TEXT NOT NULL,
+    textsearchable_index_col tsvector GENERATED ALWAYS AS (
+        to_tsvector('english', chunk_text)
+    ) STORED,
+    PRIMARY KEY (project_id, search_index_id, search_index_type, chunk_index),
+    FOREIGN KEY (search_index_id, search_index_type, project_id)
+        REFERENCES search_index(id, type, project_id)
+        ON UPDATE CASCADE ON DELETE CASCADE
+)
+""")
+
+CREATE_POSTGRES_SEARCH_INDEX_FTS_CHUNKS_INDEX = DDL("""
+CREATE INDEX IF NOT EXISTS idx_search_index_fts_chunks_fts
+ON search_index_fts_chunks USING gin(textsearchable_index_col)
+""")
+
 CREATE_POSTGRES_SEARCH_INDEX_METADATA = DDL("""
 CREATE INDEX IF NOT EXISTS idx_search_index_metadata_gin ON search_index USING gin(metadata jsonb_path_ops)
 """)