Skip to content

v4.3.2: Agent-native retrieval and reliable PDF ingestion

Choose a tag to compare

View all tags
@HabermannR HabermannR released this 27 Sep 21:04
· 1 commit to master since this release
v4.3.2
00bb559

Nexidion 4.3.2

This release makes Nexidion's content easier for AI agents to read accurately,
makes PDF ingestion reliable on small hosts, and fixes several privacy and
data-integrity bugs.

Agent-native retrieval

  • Full-text search can return verbatim snippets (snippets=true): exact passages
    from the current version with character offsets, heading path, version, the
    fields that matched, and a summary_only flag for hits found only through an
    AI summary. Search can be limited to a subtree, content kind, or authority.
  • Node-scoped assets: GET /nodes/<id>/assets lists the images a node embeds;
    GET /nodes/<id>/assets/<asset> serves one (original or downscaled preview)
    under that node's access policy.
  • Context bundles: GET /nodes/<id>/context returns a node's parent, children,
    outlinks, and backlinks, breadth-first and bounded, with relation, distance,
    link context, version, and summary status for every item. GET /nodes/context-search seeds the same expansion from search results. Hidden
    nodes are omitted without trace.

PDF ingestion

  • Optional background ingestion (background=true, used by the UI): the upload
    returns immediately and a worker thread in the app imports the file. The task
    runner can share the queue (NEXIDION_INGEST_WORKER=off leaves it to the runner).
  • Runs record a heartbeat; runs abandoned by a stopped worker are retried once and
    then failed with an explanation.
  • Each PDF is extracted once instead of twice.
  • Chapter mode no longer drops pages before the first outline entry, and tolerates
    out-of-range or duplicate outline entries.
  • gunicorn.conf.py: gthread worker with a 600-second timeout. The previous
    defaults (one sync worker, 30 seconds) killed long PDF uploads mid-import and
    blocked every other request meanwhile.

Fixes

  • Copying a node out of an AI-invisible or quarantined branch no longer makes the
    copy AI-readable.
  • Vault export and import preserve access policies, content kind, authority,
    language, tags, and metadata; export timestamps are valid ISO 8601.
  • AI actors can no longer fetch an asset by direct URL unless a node they may read
    embeds it.
  • "Generate summary" no longer fails with an internal error.
  • Deleting a user who uploaded files or requested summaries, imports, or tasks no
    longer fails.
  • Task batches are all-or-nothing.
  • Worker-completed summaries and PDF imports refresh the cached vault tree.
  • The database driver is named explicitly (postgresql+psycopg2). SQLAlchemy 2.1
    maps a bare postgresql:// URL to psycopg 3, which is not installed, so fresh
    images failed at startup.
  • Malformed vault imports are rejected with a clear error; request bodies are
    capped at 128 MiB (NEXIDION_MAX_UPLOAD_MB).

Database migration

Migration b7d3e2a91c40 adds lifecycle columns to ingestion_runs (artifact,
stored request, attempts, start time, heartbeat). Runs still marked processing
more than 15 minutes after creation are marked failed as interrupted. The
downgrade removes the columns.

Compatibility

Nexidion MCP 1.3.0 is the corresponding connector release; its new tools need the
4.3.2 endpoints. Deploy the application before the connector.

Assets 2
Loading