@babav/knowledge-core-client 0.56.0 → 0.58.0

package/dist/index.d.ts

@@ -648,6 +648,52 @@ export interface RerankerComparison {
     eval_version: string;
     heads: RerankerHead[];
 }
+/** One visual-pipeline arm measured in the visual benchmark (yield + independent graded quality). */
+export interface VisualArm {
+    arm: string;
+    generation_model: string | null;
+    concept_model: string | null;
+    concept_mode: string | null;
+    visuals_per_query: number | null;
+    zero_visual_rate: number | null;
+    dropped: number | null;
+    avg_quality: number | null;
+    n_visuals_judged: number | null;
+    quality_judge: string | null;
+    quality_caveat: string | null;
+    avg_cost_per_query_usd: number | null;
+    p50_ms: number | null;
+    n_questions: number | null;
+}
+/** Visual-benchmark leaderboard (GET /v1/benchmark/visuals) — per visual-pipeline arm: yield,
+ *  independent graded quality, cost, latency. `arms` is empty if no visual benchmark has run. */
+export interface VisualComparison {
+    eval_version: string;
+    note: string | null;
+    arms: VisualArm[];
+}
+/** One citation configuration measured in the citation benchmark. */
+export interface CitationArm {
+    arm: string;
+    generation_model: string | null;
+    citation_model: string | null;
+    mechanism: string | null;
+    precision: number | null;
+    avg_citations_per_answer: number | null;
+    zero_citation_rate: number | null;
+    avg_cost_per_query_usd: number | null;
+    p50_ms: number | null;
+    n_questions: number | null;
+    note: string | null;
+}
+/** Citation-benchmark leaderboard (GET /v1/benchmark/citations) — per configuration: citation
+ *  precision + yield + cost + latency. Native citations (~1.0 by construction) vs post-hoc (lower
+ *  ceiling); each arm's `mechanism` says which. `arms` empty if no citation benchmark has run. */
+export interface CitationComparison {
+    eval_version: string;
+    note: string | null;
+    arms: CitationArm[];
+}
 export interface Tenant {
     id: UUID;
     name: string;
@@ -1122,6 +1168,22 @@ export declare class KnowledgeCoreClient extends HttpBase {
         rerankers: (q?: {
             eval_version?: string;
         }) => Promise<RerankerComparison>;
+        /** Visual-benchmark leaderboard — per visual-pipeline arm (e.g. all-Flash vs all-Opus): yield
+         *  (visuals/query, zero-visual rate, drops), independent graded quality, cost, latency. Read-only,
+         *  tenant-anonymous. `arms` is empty when no visual benchmark has run. NOTE: quality is a Gemini
+         *  vision judge — a Gemini-family concept model (Flash) may be self-flattered, so cross-family arms
+         *  (Opus) are the cleaner quality figure (see each arm's `quality_caveat`). */
+        visuals: (q?: {
+            eval_version?: string;
+        }) => Promise<VisualComparison>;
+        /** Citation-benchmark leaderboard — per configuration (arm): citation precision, yield
+         *  (citations/answer, zero-citation rate), cost, latency. NATIVE-citation arms (Anthropic gen)
+         *  score ~1.0 by construction (verbatim inline quotes); POST-HOC arms (non-native gen + an
+         *  attributor) have a much lower ceiling regardless of attributor — see each arm's `mechanism`
+         *  and `note`. Read-only, tenant-anonymous. `arms` empty when no citation benchmark has run. */
+        citations: (q?: {
+            eval_version?: string;
+        }) => Promise<CitationComparison>;
         /** Kick the sweep as the BENCHMARK TENANT (this client's own key — no admin key). The benchmark
          *  tenant owns the bench corpora + eval-set; the sweep FULL-scores the sensible config set and
          *  populates the leaderboard. Runs in the background — returns 202 immediately. */

package/dist/index.js

@@ -487,6 +487,18 @@ export class KnowledgeCoreClient extends HttpBase {
          *  with quality (full-150 composite_q + delta) and measured end-to-end latency/cost. Read-only,
          *  tenant-anonymous. `heads` is empty when no reranker benchmark has run for the eval version. */
         rerankers: (q) => this.request("GET", "/v1/benchmark/rerankers", { query: q }),
+        /** Visual-benchmark leaderboard — per visual-pipeline arm (e.g. all-Flash vs all-Opus): yield
+         *  (visuals/query, zero-visual rate, drops), independent graded quality, cost, latency. Read-only,
+         *  tenant-anonymous. `arms` is empty when no visual benchmark has run. NOTE: quality is a Gemini
+         *  vision judge — a Gemini-family concept model (Flash) may be self-flattered, so cross-family arms
+         *  (Opus) are the cleaner quality figure (see each arm's `quality_caveat`). */
+        visuals: (q) => this.request("GET", "/v1/benchmark/visuals", { query: q }),
+        /** Citation-benchmark leaderboard — per configuration (arm): citation precision, yield
+         *  (citations/answer, zero-citation rate), cost, latency. NATIVE-citation arms (Anthropic gen)
+         *  score ~1.0 by construction (verbatim inline quotes); POST-HOC arms (non-native gen + an
+         *  attributor) have a much lower ceiling regardless of attributor — see each arm's `mechanism`
+         *  and `note`. Read-only, tenant-anonymous. `arms` empty when no citation benchmark has run. */
+        citations: (q) => this.request("GET", "/v1/benchmark/citations", { query: q }),
         /** Kick the sweep as the BENCHMARK TENANT (this client's own key — no admin key). The benchmark
          *  tenant owns the bench corpora + eval-set; the sweep FULL-scores the sensible config set and
          *  populates the leaderboard. Runs in the background — returns 202 immediately. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@babav/knowledge-core-client",
-  "version": "0.56.0",
+  "version": "0.58.0",
   "description": "TypeScript client for the Babav Knowledge Core API (Deno + Node 18+, zero deps). Includes the babav.visual grammar TYPES at the ./visual subpath (types only; all visual rendering is server-side).",
   "type": "module",
   "main": "./dist/index.js",

package/src/index.ts

@@ -613,6 +613,52 @@ export interface RerankerComparison {
   eval_version: string;
   heads: RerankerHead[];
 }
+/** One visual-pipeline arm measured in the visual benchmark (yield + independent graded quality). */
+export interface VisualArm {
+  arm: string;                               // e.g. "all_flash" | "all_opus"
+  generation_model: string | null;
+  concept_model: string | null;
+  concept_mode: string | null;
+  visuals_per_query: number | null;          // yield: mean shipped visuals per query
+  zero_visual_rate: number | null;           // fraction of queries that shipped no visual
+  dropped: number | null;
+  avg_quality: number | null;                // 0..1, independent vision judge
+  n_visuals_judged: number | null;
+  quality_judge: string | null;
+  quality_caveat: string | null;             // e.g. same-family-judge self-preference note
+  avg_cost_per_query_usd: number | null;
+  p50_ms: number | null;
+  n_questions: number | null;
+}
+/** Visual-benchmark leaderboard (GET /v1/benchmark/visuals) — per visual-pipeline arm: yield,
+ *  independent graded quality, cost, latency. `arms` is empty if no visual benchmark has run. */
+export interface VisualComparison {
+  eval_version: string;
+  note: string | null;
+  arms: VisualArm[];
+}
+/** One citation configuration measured in the citation benchmark. */
+export interface CitationArm {
+  arm: string;                               // e.g. "opus_native" | "flash_self" | "flash_opus"
+  generation_model: string | null;
+  citation_model: string | null;            // post-hoc attributor; null/native for Anthropic gen
+  mechanism: string | null;                  // "native" (inline, verbatim) | "posthoc" (attributor pass)
+  precision: number | null;                  // 0..1 — cited source actually supports the claim
+  avg_citations_per_answer: number | null;
+  zero_citation_rate: number | null;         // fraction of answers with no citation
+  avg_cost_per_query_usd: number | null;
+  p50_ms: number | null;
+  n_questions: number | null;
+  note: string | null;
+}
+/** Citation-benchmark leaderboard (GET /v1/benchmark/citations) — per configuration: citation
+ *  precision + yield + cost + latency. Native citations (~1.0 by construction) vs post-hoc (lower
+ *  ceiling); each arm's `mechanism` says which. `arms` empty if no citation benchmark has run. */
+export interface CitationComparison {
+  eval_version: string;
+  note: string | null;
+  arms: CitationArm[];
+}
 
 export interface Tenant {
   id: UUID;
@@ -1259,6 +1305,20 @@ export class KnowledgeCoreClient extends HttpBase {
      *  tenant-anonymous. `heads` is empty when no reranker benchmark has run for the eval version. */
     rerankers: (q?: { eval_version?: string }) =>
       this.request<RerankerComparison>("GET", "/v1/benchmark/rerankers", { query: q }),
+    /** Visual-benchmark leaderboard — per visual-pipeline arm (e.g. all-Flash vs all-Opus): yield
+     *  (visuals/query, zero-visual rate, drops), independent graded quality, cost, latency. Read-only,
+     *  tenant-anonymous. `arms` is empty when no visual benchmark has run. NOTE: quality is a Gemini
+     *  vision judge — a Gemini-family concept model (Flash) may be self-flattered, so cross-family arms
+     *  (Opus) are the cleaner quality figure (see each arm's `quality_caveat`). */
+    visuals: (q?: { eval_version?: string }) =>
+      this.request<VisualComparison>("GET", "/v1/benchmark/visuals", { query: q }),
+    /** Citation-benchmark leaderboard — per configuration (arm): citation precision, yield
+     *  (citations/answer, zero-citation rate), cost, latency. NATIVE-citation arms (Anthropic gen)
+     *  score ~1.0 by construction (verbatim inline quotes); POST-HOC arms (non-native gen + an
+     *  attributor) have a much lower ceiling regardless of attributor — see each arm's `mechanism`
+     *  and `note`. Read-only, tenant-anonymous. `arms` empty when no citation benchmark has run. */
+    citations: (q?: { eval_version?: string }) =>
+      this.request<CitationComparison>("GET", "/v1/benchmark/citations", { query: q }),
     /** Kick the sweep as the BENCHMARK TENANT (this client's own key — no admin key). The benchmark
      *  tenant owns the bench corpora + eval-set; the sweep FULL-scores the sensible config set and
      *  populates the leaderboard. Runs in the background — returns 202 immediately. */