@lunora/bindings 1.0.0-alpha.71 → 1.0.0-alpha.72

package/dist/packem_shared/createContextVectors-fqM-VsVd.mjs

@@ -0,0 +1,6 @@
+import{c as w,U as v}from"./concurrent-vRmSvRpF.mjs";const M=(s,r)=>{if(!r.includes("."))return s[r];let d=s;for(const n of r.split(".")){if(d===null||typeof d!="object"||Array.isArray(d))return;d=d[n]}return d},A=(s,r)=>{const d=r?.namespace,n=new Set(r?.shardedIndexNames),c=(a,i)=>{if(i!==void 0)return i;if(n.has(a)){if(d!==void 0)return d;throw new Error(`@lunora/bindings/vectors: index "${a}" belongs to a sharded table, but this DO instance has no shard key (it is the root/default DO) and no explicit namespace was given. A namespace-less operation here would reach every tenant's vectors — Vectorize indexes are account-global. Pass an explicit namespace, or issue this call from the sharded DO instance that owns the tenant.`)}},e=r?.deferAfterCommit,l=async(a,i,m)=>{await s.upsert(a,{embed:i.embed,id:i.id,input:i.input,metadata:i.metadata,namespace:m})},f=async(a,i)=>{await l(a,i,c(a,i.namespace))},o=e===void 0?f:async(a,i)=>{const m=c(a,i.namespace);await e(async()=>l(a,i,m))},t=async(a,i,m)=>{const u=await s.getByIds(a,i);return m===void 0?u:u.filter(h=>h.namespace===m)};return{deleteByIds:async(a,i,m)=>{const u=c(a,m);if(u===void 0){await s.deleteByIds(a,i);return}const h=await t(a,i,u);h.length!==0&&await s.deleteByIds(a,h.map(p=>p.id))},getByIds:async(a,i,m)=>{const u=c(a,m);return(await t(a,i,u)).map(p=>({id:p.id,metadata:p.metadata,namespace:p.namespace,values:[...p.values]}))},query:async(a,i)=>{const m=await s.query(a,{embed:i.embed,filter:i.filter,input:i.input,namespace:c(a,i.namespace),returnMetadata:i.returnMetadata??"indexed",topK:i.topK,vector:i.vector});return{count:m.count,matches:m.matches.map(u=>({id:u.id,metadata:u.metadata,score:u.score}))}},upsert:o,upsertNow:f}},y=new Set,b=s=>{y.has(s)||(y.add(s),console.warn(`[@lunora/bindings/vectors] index "${s}" syncs vectors without a namespace — in a
+multi-tenant/sharded app this exposes one tenant's vectors (and any captured
+metadata) to every other tenant, since Vectorize indexes are account-global.
+Pass \`namespace\` (the shard/tenant key) on both write and query — query-side
+namespace filtering is mandatory for multi-tenant apps. Single-tenant apps that
+legitimately have no tenant key suppress this via { allowSharedNamespace: true }.`))},S=(s,r)=>{const d={};for(const n of r)n in s&&(d[n]=s[n]);return d},g=(s,r)=>{const d=s.tables[r.table],n=d?.vectorIndexes??[],c=Object.entries(s.vectorIndexes).filter(([,t])=>t.table===r.table);if(n.length===0&&c.length===0)return;const e=d?.softDeleteMode?.field,l=e!==void 0&&r.doc?.[e]!==void 0&&r.doc[e]!==null;if(r.op==="delete"||l)return{deletes:[...n.map(t=>t.name),...c.map(([t])=>t)],upserts:[]};const f=r.doc;if(!f)return;const o={deletes:[],upserts:[]};for(const t of n){const a=M(f,t.field);if(a==null)o.deletes.push(t.name);else if(typeof a=="string")o.upserts.push({embed:t.embed,input:a,metadata:t.metadata?S(f,t.metadata):void 0,name:t.name});else throw new TypeError(`@lunora/bindings/vectors: inline index "${t.name}" expects a string source at "${t.field}" on table "${r.table}" (got ${typeof a}); use a standalone defineVectorIndex with a select() to derive text from non-string columns`)}for(const[t,a]of c)o.upserts.push({embed:a.embed,input:a.select(f),metadata:a.metadata?.(f),name:t});return o},D=s=>{const{allowSharedNamespace:r,namespace:d,schema:n,vectors:c}=s;return async e=>{const l=g(n,e);if(!l)return;const f=[...l.deletes.map(o=>async()=>{await c.deleteByIds(o,[e.id])}),...l.upserts.map(o=>async()=>{!r&&d===void 0&&b(o.name),await c.upsertNow(o.name,{embed:o.embed,id:e.id,input:o.input,metadata:o.metadata,namespace:d})})];await w(f,v,async o=>o())}},I=1e3,x=(s,r)=>{const d=[];for(let n=0;n<s.length;n+=r)d.push(s.slice(n,n+r));return d},k=async(s,r)=>{const d=await w(s,v,async e=>{try{return{item:e,ok:!0,value:await r(e)}}catch(l){return{error:l,item:e,ok:!1}}}),n=d.flatMap(e=>e.ok?[{item:e.item,value:e.value}]:[]),c=d.flatMap(e=>e.ok?[]:[{error:e.error,item:e.item}]);if(s.length>=2&&n.length===0)throw c[0]?.error;return{failed:c,ok:n}},B=(s,r,d,n)=>{const c=new Map,e=[];for(const{doc:l,id:f}of d){let o;try{o=g(s,{doc:l,id:f,op:"update",table:r})}catch(t){n.set(f,t);continue}for(const t of o?.deletes??[])c.set(t,[...c.get(t)??[],f]);for(const t of o?.upserts??[])e.push({id:f,upsert:t})}return{deletes:c,pending:e}},C=async(s,r,d,n)=>{const{allowSharedNamespace:c,namespace:e,upsertMany:l,vectors:f}=s;!c&&e===void 0&&b(r);const o=d.map(({id:t,upsert:a,values:i})=>({embed:()=>i,id:t,input:a.input,metadata:a.metadata,namespace:e}));for(const t of x(o,I))try{await l(r,t)}catch{const a=await k(t,async i=>f.upsertNow(r,i));for(const{error:i,item:m}of a.failed)n.set(m.id,i)}},E=s=>async(r,d)=>{const n=new Map,{deletes:c,pending:e}=B(s.schema,r,d,n),l=await k(e,async({upsert:o})=>o.embed(o.input)),f=new Map;for(const{error:o,item:t}of l.failed)n.set(t.id,o);for(const{item:o,value:t}of l.ok)f.set(o.upsert.name,[...f.get(o.upsert.name)??[],{...o,values:t}]);for(const[o,t]of c)for(const a of x(t,I))await s.vectors.deleteByIds(o,a);for(const[o,t]of f)await C(s,o,t,n);return[...n].map(([o,t])=>({error:t,id:o}))},V=s=>{const r=new Map,d=(n,c,e)=>{const l=e===void 0?c:[...c,e];r.set(n,[...r.get(n)??[],JSON.stringify(l)])};for(const[n,c]of Object.entries(s.tables))for(const e of c.vectorIndexes??[])d(n,[e.name,e.field,e.dimensions,e.metric,e.metadata??[]],e.model);for(const[n,c]of Object.entries(s.vectorIndexes))d(c.table,[n,"(select)",c.dimensions,c.metric],c.model);return[...r].map(([n,c])=>({profile:c.toSorted((e,l)=>e.localeCompare(l)).join("|"),table:n}))};export{A as createContextVectors,E as createVectorBackfillSync,D as createVectorSyncHook,V as vectorBackfillTargets};

package/dist/vectors/index.d.mts

@@ -219,6 +219,8 @@ interface TableVectorIndexLike {
   field: string;
   metadata?: ReadonlyArray<string>;
   metric?: string;
+  /** Declared identifier of what `embed` produces; part of the backfill fingerprint. */
+  model?: string;
   name: string;
 }
 interface TableDefinitionLike {
@@ -234,6 +236,8 @@ interface VectorIndexDefinitionLike {
   embed: VectorEmbedderLike;
   metadata?: (row: Record<string, unknown>) => Record<string, unknown>;
   metric?: string;
+  /** Declared identifier of what `embed` produces; part of the backfill fingerprint. */
+  model?: string;
   select: (row: Record<string, unknown>) => string;
   table: string;
 }
@@ -357,11 +361,15 @@ declare const createVectorBackfillSync: (options: BackfillSyncOptions) => Vector
  * backfill, which re-walks a table whose fingerprint changed.
  *
  * Covers what the schema can see: index names, the inline source field,
- * dimensions, metric and inline metadata fields. A function (`embed`, a Shape B
- * `select`/`metadata`) has no stable identity to fingerprint — its source text
- * changes with unrelated rebuilds of the bundle, which would re-embed whole
- * tables for nothing — so changing one is announced by calling the backfill with
- * `restart: true`.
+ * dimensions, metric, inline metadata fields and the declared `model`. A
+ * function (`embed`, a Shape B `select`/`metadata`) has no stable identity to
+ * fingerprint — its source text changes with unrelated rebuilds of the bundle,
+ * which would re-embed whole tables for nothing — so the declared `model` string
+ * stands in for `embed`, and any other change is announced by calling the
+ * backfill with `restart: true`.
+ *
+ * `model` joins a descriptor only when declared, so an index without one keeps
+ * the fingerprint it was recorded under and is not re-embedded for it.
  */
 declare const vectorBackfillTargets: (schema: SchemaLike) => {
   profile: string;

package/dist/vectors/index.d.ts

@@ -219,6 +219,8 @@ interface TableVectorIndexLike {
   field: string;
   metadata?: ReadonlyArray<string>;
   metric?: string;
+  /** Declared identifier of what `embed` produces; part of the backfill fingerprint. */
+  model?: string;
   name: string;
 }
 interface TableDefinitionLike {
@@ -234,6 +236,8 @@ interface VectorIndexDefinitionLike {
   embed: VectorEmbedderLike;
   metadata?: (row: Record<string, unknown>) => Record<string, unknown>;
   metric?: string;
+  /** Declared identifier of what `embed` produces; part of the backfill fingerprint. */
+  model?: string;
   select: (row: Record<string, unknown>) => string;
   table: string;
 }
@@ -357,11 +361,15 @@ declare const createVectorBackfillSync: (options: BackfillSyncOptions) => Vector
  * backfill, which re-walks a table whose fingerprint changed.
  *
  * Covers what the schema can see: index names, the inline source field,
- * dimensions, metric and inline metadata fields. A function (`embed`, a Shape B
- * `select`/`metadata`) has no stable identity to fingerprint — its source text
- * changes with unrelated rebuilds of the bundle, which would re-embed whole
- * tables for nothing — so changing one is announced by calling the backfill with
- * `restart: true`.
+ * dimensions, metric, inline metadata fields and the declared `model`. A
+ * function (`embed`, a Shape B `select`/`metadata`) has no stable identity to
+ * fingerprint — its source text changes with unrelated rebuilds of the bundle,
+ * which would re-embed whole tables for nothing — so the declared `model` string
+ * stands in for `embed`, and any other change is announced by calling the
+ * backfill with `restart: true`.
+ *
+ * `model` joins a descriptor only when declared, so an index without one keeps
+ * the fingerprint it was recorded under and is not re-embedded for it.
  */
 declare const vectorBackfillTargets: (schema: SchemaLike) => {
   profile: string;

package/dist/vectors/index.mjs

@@ -1 +1 @@
-import{createContextVectors as t,createVectorBackfillSync as o,createVectorSyncHook as c,vectorBackfillTargets as a}from"../packem_shared/createContextVectors-Bz5Me4s6.mjs";import{createVectorAdminIntrospector as l}from"../packem_shared/createVectorAdminIntrospector-Ct8v6PxJ.mjs";import{default as s}from"../packem_shared/createVectors-Dzv0ilKE.mjs";export{t as createContextVectors,l as createVectorAdminIntrospector,o as createVectorBackfillSync,c as createVectorSyncHook,s as createVectors,a as vectorBackfillTargets};
+import{createContextVectors as t,createVectorBackfillSync as o,createVectorSyncHook as c,vectorBackfillTargets as a}from"../packem_shared/createContextVectors-fqM-VsVd.mjs";import{createVectorAdminIntrospector as l}from"../packem_shared/createVectorAdminIntrospector-Ct8v6PxJ.mjs";import{default as s}from"../packem_shared/createVectors-Dzv0ilKE.mjs";export{t as createContextVectors,l as createVectorAdminIntrospector,o as createVectorBackfillSync,c as createVectorSyncHook,s as createVectors,a as vectorBackfillTargets};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lunora/bindings",
-  "version": "1.0.0-alpha.71",
+  "version": "1.0.0-alpha.72",
   "description": "Lightweight Cloudflare binding helpers for Lunora — ctx.kv, ctx.images, ctx.analytics, ctx.pipelines, ctx.vectors, ctx.r2sql — one install, per-binding subpaths",
   "keywords": [
     "analytics",
@@ -65,7 +65,7 @@
   },
   "dependencies": {
     "@lunora/errors": "1.0.0-alpha.41",
-    "@lunora/platform": "1.0.0-alpha.36"
+    "@lunora/platform": "1.0.0-alpha.37"
   },
   "engines": {
     "node": "^22.15.0 || >=24.11.0"

package/dist/packem_shared/createContextVectors-Bz5Me4s6.mjs

@@ -1,6 +0,0 @@
-import{c as w,U as v}from"./concurrent-vRmSvRpF.mjs";const M=(s,r)=>{if(!r.includes("."))return s[r];let i=s;for(const a of r.split(".")){if(i===null||typeof i!="object"||Array.isArray(i))return;i=i[a]}return i},A=(s,r)=>{const i=r?.namespace,a=new Set(r?.shardedIndexNames),d=(t,c)=>{if(c!==void 0)return c;if(a.has(t)){if(i!==void 0)return i;throw new Error(`@lunora/bindings/vectors: index "${t}" belongs to a sharded table, but this DO instance has no shard key (it is the root/default DO) and no explicit namespace was given. A namespace-less operation here would reach every tenant's vectors — Vectorize indexes are account-global. Pass an explicit namespace, or issue this call from the sharded DO instance that owns the tenant.`)}},n=r?.deferAfterCommit,f=async(t,c,u)=>{await s.upsert(t,{embed:c.embed,id:c.id,input:c.input,metadata:c.metadata,namespace:u})},l=async(t,c)=>{await f(t,c,d(t,c.namespace))},o=n===void 0?l:async(t,c)=>{const u=d(t,c.namespace);await n(async()=>f(t,c,u))},e=async(t,c,u)=>{const m=await s.getByIds(t,c);return u===void 0?m:m.filter(h=>h.namespace===u)};return{deleteByIds:async(t,c,u)=>{const m=d(t,u);if(m===void 0){await s.deleteByIds(t,c);return}const h=await e(t,c,m);h.length!==0&&await s.deleteByIds(t,h.map(p=>p.id))},getByIds:async(t,c,u)=>{const m=d(t,u);return(await e(t,c,m)).map(p=>({id:p.id,metadata:p.metadata,namespace:p.namespace,values:[...p.values]}))},query:async(t,c)=>{const u=await s.query(t,{embed:c.embed,filter:c.filter,input:c.input,namespace:d(t,c.namespace),returnMetadata:c.returnMetadata??"indexed",topK:c.topK,vector:c.vector});return{count:u.count,matches:u.matches.map(m=>({id:m.id,metadata:m.metadata,score:m.score}))}},upsert:o,upsertNow:l}},y=new Set,b=s=>{y.has(s)||(y.add(s),console.warn(`[@lunora/bindings/vectors] index "${s}" syncs vectors without a namespace — in a
-multi-tenant/sharded app this exposes one tenant's vectors (and any captured
-metadata) to every other tenant, since Vectorize indexes are account-global.
-Pass \`namespace\` (the shard/tenant key) on both write and query — query-side
-namespace filtering is mandatory for multi-tenant apps. Single-tenant apps that
-legitimately have no tenant key suppress this via { allowSharedNamespace: true }.`))},S=(s,r)=>{const i={};for(const a of r)a in s&&(i[a]=s[a]);return i},g=(s,r)=>{const i=s.tables[r.table],a=i?.vectorIndexes??[],d=Object.entries(s.vectorIndexes).filter(([,e])=>e.table===r.table);if(a.length===0&&d.length===0)return;const n=i?.softDeleteMode?.field,f=n!==void 0&&r.doc?.[n]!==void 0&&r.doc[n]!==null;if(r.op==="delete"||f)return{deletes:[...a.map(e=>e.name),...d.map(([e])=>e)],upserts:[]};const l=r.doc;if(!l)return;const o={deletes:[],upserts:[]};for(const e of a){const t=M(l,e.field);if(t==null)o.deletes.push(e.name);else if(typeof t=="string")o.upserts.push({embed:e.embed,input:t,metadata:e.metadata?S(l,e.metadata):void 0,name:e.name});else throw new TypeError(`@lunora/bindings/vectors: inline index "${e.name}" expects a string source at "${e.field}" on table "${r.table}" (got ${typeof t}); use a standalone defineVectorIndex with a select() to derive text from non-string columns`)}for(const[e,t]of d)o.upserts.push({embed:t.embed,input:t.select(l),metadata:t.metadata?.(l),name:e});return o},D=s=>{const{allowSharedNamespace:r,namespace:i,schema:a,vectors:d}=s;return async n=>{const f=g(a,n);if(!f)return;const l=[...f.deletes.map(o=>async()=>{await d.deleteByIds(o,[n.id])}),...f.upserts.map(o=>async()=>{!r&&i===void 0&&b(o.name),await d.upsertNow(o.name,{embed:o.embed,id:n.id,input:o.input,metadata:o.metadata,namespace:i})})];await w(l,v,async o=>o())}},I=1e3,x=(s,r)=>{const i=[];for(let a=0;a<s.length;a+=r)i.push(s.slice(a,a+r));return i},k=async(s,r)=>{const i=await w(s,v,async n=>{try{return{item:n,ok:!0,value:await r(n)}}catch(f){return{error:f,item:n,ok:!1}}}),a=i.flatMap(n=>n.ok?[{item:n.item,value:n.value}]:[]),d=i.flatMap(n=>n.ok?[]:[{error:n.error,item:n.item}]);if(s.length>=2&&a.length===0)throw d[0]?.error;return{failed:d,ok:a}},B=(s,r,i,a)=>{const d=new Map,n=[];for(const{doc:f,id:l}of i){let o;try{o=g(s,{doc:f,id:l,op:"update",table:r})}catch(e){a.set(l,e);continue}for(const e of o?.deletes??[])d.set(e,[...d.get(e)??[],l]);for(const e of o?.upserts??[])n.push({id:l,upsert:e})}return{deletes:d,pending:n}},C=async(s,r,i,a)=>{const{allowSharedNamespace:d,namespace:n,upsertMany:f,vectors:l}=s;!d&&n===void 0&&b(r);const o=i.map(({id:e,upsert:t,values:c})=>({embed:()=>c,id:e,input:t.input,metadata:t.metadata,namespace:n}));for(const e of x(o,I))try{await f(r,e)}catch{const t=await k(e,async c=>l.upsertNow(r,c));for(const{error:c,item:u}of t.failed)a.set(u.id,c)}},E=s=>async(r,i)=>{const a=new Map,{deletes:d,pending:n}=B(s.schema,r,i,a),f=await k(n,async({upsert:o})=>o.embed(o.input)),l=new Map;for(const{error:o,item:e}of f.failed)a.set(e.id,o);for(const{item:o,value:e}of f.ok)l.set(o.upsert.name,[...l.get(o.upsert.name)??[],{...o,values:e}]);for(const[o,e]of d)for(const t of x(e,I))await s.vectors.deleteByIds(o,t);for(const[o,e]of l)await C(s,o,e,a);return[...a].map(([o,e])=>({error:e,id:o}))},V=s=>{const r=new Map,i=(a,d)=>{r.set(a,[...r.get(a)??[],JSON.stringify(d)])};for(const[a,d]of Object.entries(s.tables))for(const n of d.vectorIndexes??[])i(a,[n.name,n.field,n.dimensions,n.metric,n.metadata??[]]);for(const[a,d]of Object.entries(s.vectorIndexes))i(d.table,[a,"(select)",d.dimensions,d.metric]);return[...r].map(([a,d])=>({profile:d.toSorted((n,f)=>n.localeCompare(f)).join("|"),table:a}))};export{A as createContextVectors,E as createVectorBackfillSync,D as createVectorSyncHook,V as vectorBackfillTargets};