service-content — entity extraction and knowledge-graph host

An organization's files hold its knowledge but do not surface it. An email archive, a folder of contracts, a store of PDFs — each is searchable, and none of it knows who relates to whom or what the organization's own terms mean. service-content is the platform's entity-extraction and knowledge-graph write service: it turns raw document payloads into named entities and relationships, and it owns the reference taxonomies those entities are classified against.

Extraction runs through three tiers, cheapest first

When a payload arrives, service-content tries the fastest extraction method first and only escalates when it has to:

1. Tier 0 — GLiNER. A direct HTTP call to a local GLiNER model (port 9085), with no call to the AI request router (the Doorman) involved. Most documents are handled here.
2. Tier A fallback — OLMo, via the Doorman. Runs when GLiNER is unreachable, when GLiNER finds nothing (to catch its blind spots), or when the payload is structured CSV data GLiNER's natural-language model can't parse. Every Tier 0 pass also queues an asynchronous Tier A run in the background. The two results become a training pair — GLiNER's extractive output as the preferred answer, OLMo's as the comparison — used to keep improving Tier A's own extraction quality over time. This is the platform's real self-improvement loop for this pipeline: a concrete training-data mechanism, not an organically growing glossary.
3. Tier B — OLMo 32B, via the Doorman's /v1/extract endpoint. The heaviest tier, used when the faster tiers are unavailable or insufficient.

A backpressure check against the Doorman's queue depth can defer a document rather than pile onto an already-loaded pipeline.

Every automated graph write is held for a human verdict

service-content never writes extracted entities straight into the knowledge graph. Every write — from either tier — passes through the same checkpoint. It captures the write first, then promotes it only after a human signs off:

1. Capture. The write is recorded to a durable, on-disk JSONL file instead of touching the graph. It survives a process restart, and a human can review the full pending queue at any time.
2. Verify. A reviewer submits an SSH-signed verdict. The signature is checked against the workspace's allowed_signers file under a dedicated namespace, so a verdict signature from this system can never be replayed against an unrelated one.
3. Promote on accept. Only after a verdict is signed does the write actually reach the graph.
4. Discard on reject. The pending record is kept — never deleted — with the rejection verdict attached, for audit.

This satisfies the platform's SYS-ADR-07 (no structured data through AI) and SYS-ADR-19 (no automated AI publishing to verified ledgers) commitments directly: the extraction pipeline can run unattended, but nothing it produces becomes part of the graph of record without an explicit, signed human decision.

The reference taxonomies it hosts

Separately from extraction, service-content owns a set of CSV taxonomy files — the Chart of Accounts and eleven archetypes among them, plus domain vocabularies, glossaries, and topic/guide indexes for each wiki. These load into the graph as static reference entities through a small admin API, distinct from the entities the extraction pipeline produces from real documents.

See also

• Archetypes and chart of accounts — the two reference taxonomies service-content loads into the knowledge graph
• Verification surveyor — where a human applies the archetype label that service-content's taxonomy makes available
• service-extraction — the DataGraph ingestion pipeline — the deterministic identity layer entities receive once extracted
• Console input application — the F12 gate where an operator reviews a pending automated write
• Query the DataGraph — step-by-step guide: search named entities, navigate relationships, and handle Tier A outages
• Export structured data from the platform — step-by-step guide: four export paths including DataGraph queries and wiki Markdown