@vertesia/common 1.5.0-dev.20260725.083715Z → 1.5.0-dev.20260804.124748Z

package/src/store/editing-policy-schema.generated.ts

@@ -0,0 +1,30 @@
+// GENERATED FILE — DO NOT EDIT.
+//
+// Written by `scripts/gen-intake-policy-schema.ts` from `ContentTypeEditingPolicySchema` in
+// `../api-schemas/store.ts`, through the same adapter that emits the OpenAPI components. Edit the Zod
+// schema and re-run `pnpm run gen:schemas`; `store.contract.test.ts` fails if this drifts from the
+// canonical component, and fails too if it accepts a value the Zod schema rejects.
+//
+// It exists so the Studio policy editors and the server validators get a self-contained JSON
+// Schema without importing Zod: the package root exports plain data, and `zod` stays out of every
+// browser bundle. The component's `$ref`s are re-rooted from `#/components/schemas/` to `#/$defs/`
+// so AJV and Monaco can compile it standalone.
+//
+// It carries the name consumers have always imported. The Zod object it was emitted from is exported
+// under the same name from `../api-schemas/store.ts`, which the package root does NOT re-export —
+// that split is what keeps zod out of the browser while the alias-provenance gate still sees the
+// `${Name}Schema` convention it requires.
+
+import type { JSONObject } from '../json.js';
+
+export const ContentTypeEditingPolicySchema: JSONObject = {
+    type: 'object',
+    properties: {
+        interaction: {
+            type: 'string',
+            description: 'Agent interaction used for new document-editing sessions. Defaults to sys:GeneralAgent.',
+        },
+    },
+    additionalProperties: false,
+    description: 'Per-content-type policy for collaborative document editing.',
+};

package/src/store/grounded-extraction.ts

@@ -1,4 +1,11 @@
-import type { InteractionExecutionConfiguration } from '../interaction.js';
+import type { z } from 'zod';
+import type {
+    GroundedAssistantResponseSchema,
+    GroundedExtractionRequestSchema,
+    GroundedExtractionResultResponseSchema,
+    GroundedExtractionVerdictSchema,
+    GroundedVerificationBreakdownSchema,
+} from '../api-schemas/document-processing.js';
 import type { WorkflowRunStatus } from './workflow.js';
 
 /**
@@ -7,7 +14,7 @@ import type { WorkflowRunStatus } from './workflow.js';
  * check; `needs_review` means a human should verify it. This reflects content
  * correctness, not how many citation boxes rendered.
  */
-export type GroundedExtractionVerdict = 'good_to_go' | 'needs_review';
+export type GroundedExtractionVerdict = z.infer<typeof GroundedExtractionVerdictSchema>;
 
 /**
  * Canonical workflow id for object-scoped grounded extraction. Every entry point
@@ -22,69 +29,7 @@ export function getGroundedExtractionWorkflowId(accountId: string, objectId: str
  * optional: with none set, the object's own content-type schema drives the
  * extraction with default models and settings.
  */
-export interface GroundedExtractionRequest {
-    /** JSON schema describing the data to extract. Takes precedence over type_ref. */
-    schema?: Record<string, unknown>;
-    /** Content type id or catalog ref whose object_schema drives the extraction. */
-    type_ref?: string;
-    /** Interaction to use. Defaults to sys:ExtractInformationGrounded. */
-    interaction_name?: string;
-    /** Maximum number of pages to process. */
-    max_pages?: number;
-    /** Run OCR on every page even when a text layer exists. */
-    force_ocr?: boolean;
-    /** Re-run OCR on pages that need it instead of restoring the stored OCR result. */
-    refresh_ocr?: boolean;
-    /** Attach clean page images for layout/semantic context; direct-vision pages also receive checkerboards. */
-    use_vision?: boolean;
-    /**
-     * A1 locate-grid cell size in PDF points for vision pages (drives both the drawn
-     * grid and cell→box resolution). Smaller = finer grid / more cells. Default: 14.
-     */
-    grid_cell_pt?: number;
-    /**
-     * How to read pages that have no digital text layer (scans / image-only pages).
-     * 'vision' (default): read them off the page image with the extraction model,
-     * skipping OCR entirely. 'ocr': legacy path — OCR those pages and block-ground on
-     * the (lossy) OCR text. Set to 'ocr' to revert to the pre-vision behavior.
-     */
-    raster_mode?: 'vision' | 'ocr';
-    /** Maximum pages per extraction call; larger documents are split into sequential windows. */
-    window_pages?: number;
-    /**
-     * Extract with an autonomous agent (views the whole document at once) instead of
-     * the deterministic windowed pipeline. Sidesteps window-boundary splits on long
-     * documents. The workflow stages the artifacts into an agent space, runs a
-     * conversation agent that writes the extraction, then folds it back.
-     */
-    agentic_extraction?: boolean;
-    /** Agent interaction for agentic_extraction. Defaults to sys:GeneralAgent. */
-    extract_agent?: string;
-    /** Update the object's properties with the extracted data. Default: true. */
-    update_properties?: boolean;
-    /** LLM execution configuration (model, environment, ...) for the main pass. */
-    config?: InteractionExecutionConfiguration;
-    /** Execution configuration used instead of `config` on hard content (scans, handwriting). */
-    hard_config?: InteractionExecutionConfiguration;
-    /** Hardness score (0..1) at or above which `hard_config` is used. Default: 0.5. */
-    hardness_threshold?: number;
-    /** Execution configuration for the post-extraction review pass. No review runs when absent. */
-    review_config?: InteractionExecutionConfiguration;
-    /** Hardness score (0..1) at or above which the review runs. Defaults to hardness_threshold. */
-    review_threshold?: number;
-    /** Review triggers when any page's citation coverage falls below this floor. Default: 0.2. */
-    coverage_review_threshold?: number;
-    /** Run the model review even when every citation was digitally verified. Requires review_config. */
-    force_review?: boolean;
-    /**
-     * Free-text operator guidance folded into the extraction prompt to steer a
-     * (re-)extraction, e.g. "part numbers are in the third column; some line items
-     * wrap onto the next row".
-     */
-    operator_instructions?: string;
-    /** Interactive assistant only: the operator's opening message for the assistant conversation. */
-    user_prompt?: string;
-}
+export type GroundedExtractionRequest = z.infer<typeof GroundedExtractionRequestSchema>;
 
 /**
  * Response from starting the interactive grounded extraction assistant. The agent
@@ -92,14 +37,7 @@ export interface GroundedExtractionRequest {
  * document into the agent space -> launch the interactive conversation); the
  * client renders the conversation with `agent_run_id`.
  */
-export interface GroundedAssistantResponse {
-    /** The AgentRun id to stream/render the conversation. */
-    agent_run_id: string;
-    /** The conversation workflow id backing the run. */
-    workflow_id: string;
-    /** The object the assistant is scoped to. */
-    object_id: string;
-}
+export type GroundedAssistantResponse = z.infer<typeof GroundedAssistantResponseSchema>;
 
 /**
  * Status of a grounded extraction workflow. Carries the doc-level verdict once the
@@ -116,39 +54,10 @@ export interface GroundedExtractionRunStatusResponse extends WorkflowRunStatus {
  * AI verified (the reviewer confirmed it against the page image — the primary
  * signal for scanned or handwritten content, which has no text layer to match).
  */
-export interface GroundedVerificationBreakdown {
-    /** Total number of cited values. */
-    total: number;
-    /** Values matched verbatim against the document's text (digital layer or OCR). */
-    digitally_verified: number;
-    /** Values the reviewer model confirmed against the page image. */
-    ai_verified: number;
-    /** Values read from the image but neither text-matched nor reviewer-confirmed. */
-    unverified: number;
-}
+export type GroundedVerificationBreakdown = z.infer<typeof GroundedVerificationBreakdownSchema>;
 
 /**
  * Completed grounded extraction result: the extracted data with its trust verdict
  * and verification breakdown, plus a download URL for the full citations artifact.
  */
-export interface GroundedExtractionResultResponse {
-    object_id: string;
-    /** The extracted data, shaped by the requested schema. */
-    data: Record<string, unknown>;
-    /** Document-level trust verdict. */
-    verdict?: GroundedExtractionVerdict;
-    /** One-sentence rationale for the verdict. */
-    verdict_reason?: string;
-    /** Mean citation confidence in [0,1]. */
-    confidence?: number;
-    /** Per-value verification breakdown. */
-    verification: GroundedVerificationBreakdown;
-    /** Review outcome, when a review pass ran. */
-    review?: {
-        assessment: 'complete' | 'issues_found';
-        summary?: string;
-        corrections_applied?: number;
-    };
-    /** Signed download URL for the full grounded-extraction.json (data + citations + boxes). */
-    result_url?: string | null;
-}
+export type GroundedExtractionResultResponse = z.infer<typeof GroundedExtractionResultResponseSchema>;

package/src/store/index.ts

@@ -6,8 +6,11 @@ export * from './common.js';
 export * from './conversation-state.js';
 export * from './doc-analyzer.js';
 export * from './dsl-workflow.js';
+// The compiled JSON Schema for ContentTypeIntakePolicy, under the name consumers import.
+export * from './editing-policy-schema.generated.js';
 export * from './grounded-extraction.js';
 export * from './hive-memory.js';
+export * from './intake-policy-schema.generated.js';
 export * from './object-types.js';
 export * from './process.js';
 export * from './process-schema.js';