Skip to content

v0.3.10

Latest

Choose a tag to compare

@noah-paige noah-paige released this 12 Nov 22:42
· 2 commits to main since this release
c91e1b0

Features

  • The Playground now shows compiled SQL/SPARQL to schema-authorized roles and hides all schema from data analysts. Admins and namespace owners, maintainers and data stewards now see the compiled SQL and SPARQL in the Playground by default; before, it was hidden from everyone. Data analysts can still run queries and use metrics, but they get only results and answers. They see no table names, SQL, ontology classes or graph queries anywhere: not in the UI, the API, or through MCP agents.
  • Stewards review business vocabulary before accepting a proposal. A class's side panel on the proposal page edits the preferred label, adds a language, adds or rejects variants and makes a variant preferred. Structured queries also find a class by its accepted variants, as the ontologies guide describes.
  • Deep Reasoning writes better SQL by default. Its SQL writer sees the original question and the documents already read, and revises a query whose result looks wrong. In one test over documents and tables in one namespace, correct answers rose from 41% to 94% in a 57-question set. These defaults apply after you redeploy, and the serve guide explains how to turn them off per request or on every deploy.

Enhancements/Bug Fixes

  • Ontology deletes must provide the ontology ID. The API refuses a delete request without an ontology ID instead of removing all of the namespace's ontology data. Scripts that cleared a namespace that way must delete each ontology by its ID, as the ontologies guide shows.
  • Metric grants limit every kind of answer. A role's allowed metrics also limit the governed metrics that structured queries, graph queries and Deep Reasoning use, so a restricted role gets no answer built from other metrics. An empty grant allows none.
  • Structured answers are easier to read and compare. Ranking questions return the name and ranked value, not only an ID, and aggregate columns keep conventional names across runs. An opt-in majority vote across up to five candidate queries picks the answer, at up to five times the cost, as the deploying guide describes.
  • Schema.org Grounding suggests fewer wrong classes. Schema.org no longer offers the 33 classes from other vocabularies that its file declares, and a tie between same-named classes from two ontologies shows No confident match for steward review instead of creating a class. If you loaded schema.org before this release, reload it to drop those classes from grounding.
  • Guardrails no longer miss text at chunk boundaries. During document ingestion, the guardrail checks each document in 10,000-character pieces, which now overlap so that text split across two pieces is checked whole. Long documents use about 11% more guardrail calls.
  • The MCP proxy uses patched MCP libraries. Download the proxy again from the console's MCP Server page and rebuild it, because a proxy downloaded earlier keeps the old versions. The control-plane agent skill moved to a different folder, which the README shows.
  • Ontology jobs report results and next steps. Constraint inference and proposal validation no longer finish without a result or fail without an error. When a re-induction finds an unchanged schema, the console says to delete the ontology and induce again, and the ontologies guide says to curate before the first induction.
  • REST queries that run out of time say so. A query that cannot finish gets the documented deadline error about a second before the gateway's 29-second limit, instead of a generic gateway timeout. Time the authorizer takes no longer counts against the query, as the serve guide describes.
  • Source scans and errors point to the right fix. When a scan cannot read some tables, the warning says to approve the readable tables and re-scan once the rest are readable. Source error messages replace S3 locations with a placeholder, so identify files by name instead.
  • Query responses can report LLM token use. It is opt-in per request and returns the input and output tokens for each model, as the serve guide describes.
  • The Explorer graph defaults to the induced ontology. Its graph view opens on the namespace's induced ontology, as the list view already did, and All ontologies also shows reference vocabularies. The Playground help says the SQL Firewall checks each role's allowed tables and denied columns, not row filters.

Operational

  • Serve-only deploys no longer change other stacks. The serve-only deploy updates only the serve stack instead of also redeploying dependent stacks with default settings, which could undo choices such as an existing VPC or search capacity. Pass it the same settings as the standard deploy every time, because deploys do not store them, as the deploying guide describes.
  • Namespaces no longer leave VKG resources behind. Deleting a namespace removes its VKG task definitions, log groups and proposal and upload files, and VKG logs keep 30 days. Accepting a proposal no longer creates VKG services in other deployments in the account and Region. Delete by hand the log groups of namespaces deleted earlier, as the deploying guide describes, and VKG services held for another deployment's namespaces.
  • Large ontologies can get longer Deep Reasoning timeouts. Opt-in settings raise the ontology-lookup and Neptune timeouts from their 6- and 8-second defaults, and another keeps the Region's existing API Gateway logging role. Pass them on every deploy, as the deploying guide describes.
  • Deploys stop at the real error. A failed dependency install stops the document preprocessing image build with its own error instead of a misleading missing-module error. The deployed MCP server refuses to start with authentication off, and the deploying guide names AgentCore Runtime microVMs as the regional prerequisite.
  • Guardrail logs name the policy, not the matched text. Blocked and masked responses log each policy and its action, so update any log filter that reads the old assessment details.
  • The architecture guide** walks through how COA scans, models, governs metrics and serves answers.**