Configure artifact formats

Configure Pinecone Nexus artifact types as markdown prose or queryable SQLite tables through the manifest API.

In Pinecone Nexus, curation writes each artifact type in one of two formats, set per type in the manifest: markdown for prose, or sqlite for rows in a queryable table. You choose the format in the manifest you send when you create or update a context.

Markdown artifacts

markdown is the default. Curation writes one prose file per artifact, embedded for semantic retrieval, so it fits knowledge that reads as text: summaries, concepts, and per-document notes. A markdown artifact type needs a name, a kind, a scope, and a description that steers what curation extracts:

JSON

{
  "name": "Concept",
  "kind": "topic",
  "scope": "corpus",
  "description": "A recurring concept or theme that appears across the corpus."
}

Omitting format defaults to markdown. See artifact kinds for the kinds a type can take.

SQLite artifacts

sqlite writes rows into the context's structured database instead of prose, so the type becomes a table the query agent reads with SQL for exact counts, filters, and joins. Set format to sqlite and declare the table's columns:

JSON

{
  "name": "Meeting Log",
  "kind": "entity",
  "scope": "corpus",
  "format": "sqlite",
  "description": "One row per meeting: the body that met, the date, the type, and who chaired.",
  "columns": [
    { "name": "body", "type": "TEXT", "description": "The board, committee, or team that met, as a short canonical name." },
    { "name": "meeting_date", "type": "TEXT", "description": "The meeting date in ISO format, YYYY-MM-DD." },
    { "name": "year", "type": "INTEGER", "description": "The calendar year of the meeting." },
    { "name": "chair", "type": "TEXT", "description": "Full name of whoever chaired the meeting." }
  ],
  "natural_key": ["body", "meeting_date"]
}

Each column has a name, a type (TEXT, INTEGER, REAL, or NUMERIC), and a description that steers what curation extracts into it.
natural_key names the columns that identify a row, so re-curating a source upserts its rows instead of duplicating them.

Because the table is queryable, a sqlite type answers questions prose can't, like "how many meetings did each chair run this year?" or "who was present at a given meeting?" To answer across tables, define a second table that shares columns, so queries can join the two:

JSON

{
  "name": "Meeting Attendance",
  "kind": "entity",
  "scope": "corpus",
  "format": "sqlite",
  "description": "One row per person per meeting, taken from the attendance roster.",
  "columns": [
    { "name": "person", "type": "TEXT", "description": "The attendee's full name, spelled the same way wherever it appears, so counts and joins line up." },
    { "name": "body", "type": "TEXT", "description": "The body that met, matching Meeting Log." },
    { "name": "meeting_date", "type": "TEXT", "description": "The meeting date in ISO format, matching Meeting Log." },
    { "name": "status", "type": "TEXT", "description": "Either 'present' or 'absent', from the roster." }
  ],
  "natural_key": ["person", "body", "meeting_date"]
}

Set the formats through the API

Put your artifact types in a context's manifest under curate.artifacts.artifact_types, then curate. The walkthrough below seeds the Concept and Meeting Log types, and you add more the same way. These steps use the data-plane base URL and session token from Authentication.

Write the manifest to the context

Send the manifest when you create the context, or update an existing one with PUT /contexts/{slug}. This example seeds both a markdown type and a SQLite table at creation:

curl

curl -fsS -X POST "$NEXUS_BASE_URL/contexts" \
  -H "Authorization: Bearer $NEXUS_TOKEN" \
  -H 'Content-Type: application/json' \
  -H 'X-Pinecone-Api-Version: 2026-07' \
  -d '{
    "slug": "board-minutes",
    "name": "Board minutes",
    "manifest": {
      "curate": {
        "artifacts": {
          "enabled": true,
          "artifact_types": [
            { "name": "Concept", "kind": "topic", "scope": "corpus",
              "description": "A recurring concept or theme across the minutes." },
            { "name": "Meeting Log", "kind": "entity", "scope": "corpus", "format": "sqlite",
              "description": "One row per meeting: body, date, and chair.",
              "columns": [
                { "name": "body", "type": "TEXT", "description": "The body that met." },
                { "name": "meeting_date", "type": "TEXT", "description": "ISO date, YYYY-MM-DD." },
                { "name": "chair", "type": "TEXT", "description": "Who chaired the meeting." }
              ],
              "natural_key": ["body", "meeting_date"] }
          ]
        }
      }
    }
  }'

Curate the context
Curation reads the sources and builds the artifacts your manifest defines. An empty body runs the first curate:
curl
```
curl -fsS -X POST "$NEXUS_BASE_URL/contexts/board-minutes/curate" \
  -H "Authorization: Bearer $NEXUS_TOKEN" \
  -H 'Content-Type: application/json' \
  -H 'X-Pinecone-Api-Version: 2026-07' \
  -d '{}'
```
After a later manifest change, pass {"force": true} to rebuild everything under the new manifest. Curation runs as a background task, so query the context once it finishes. See How curation works.

Query across both formats

A query reads both formats: the prose Concept artifacts semantically, and the Meeting Log table with SQL for exact answers.

curl

curl -fsS -X POST "$NEXUS_BASE_URL/query" \
  -H "Authorization: Bearer $NEXUS_TOKEN" \
  -H 'Content-Type: application/json' \
  -H 'X-Pinecone-Api-Version: 2026-07' \
  -d '{"scope": ["board-minutes"], "ask": "How many meetings did each chair run this year?"}' \
| jq -r '.output[].content[].text'

Output

Jordan Lee chaired 8 meetings, Priya Shah chaired 5, and Marcus Cole chaired 3.

The counts come from the Meeting Log table, so they're exact rather than estimated from prose.