@lunora/ai 1.0.0-alpha.95 → 1.0.0-alpha.97

package/dist/index.mjs

@@ -1 +1 @@
-import{default as _}from"./packem_shared/createAi-CHSAfN98.mjs";import{AI_DEFAULT_EMBEDDING_MODEL_ENV as a,AI_DEFAULT_MODEL_ENV as E,AI_GATEWAY_ACCOUNT_ID_ENV as o,AI_GATEWAY_ID_ENV as r,AI_GATEWAY_METADATA_MAX_KEYS as T,AI_GATEWAY_TAGS_ENV as I,AI_GATEWAY_TOKEN_ENV as l,buildAiGatewayMetadataFields as m,readAiGatewayEnvTags as s,resolveAiGateway as D}from"./packem_shared/AI_DEFAULT_EMBEDDING_MODEL_ENV-BEAzcoBw.mjs";import{DEFAULT_MODEL_PRICES as M,estimateModelCost as d,lookupModelPrice as N}from"./packem_shared/DEFAULT_MODEL_PRICES-Q8uxdiuV.mjs";import{embed as x,embedMany as O,generateObject as c,generateText as f,hasToolCall as p,jsonSchema as L,streamObject as V,streamText as W,tool as Y}from"ai";import{createWorkersAI as n}from"workers-ai-provider";export{a as AI_DEFAULT_EMBEDDING_MODEL_ENV,E as AI_DEFAULT_MODEL_ENV,o as AI_GATEWAY_ACCOUNT_ID_ENV,r as AI_GATEWAY_ID_ENV,T as AI_GATEWAY_METADATA_MAX_KEYS,I as AI_GATEWAY_TAGS_ENV,l as AI_GATEWAY_TOKEN_ENV,M as DEFAULT_MODEL_PRICES,m as buildAiGatewayMetadataFields,_ as createAi,n as createWorkersAI,x as embed,O as embedMany,d as estimateModelCost,c as generateObject,f as generateText,p as hasToolCall,L as jsonSchema,N as lookupModelPrice,s as readAiGatewayEnvTags,D as resolveAiGateway,V as streamObject,W as streamText,Y as tool};
+import{default as _}from"./packem_shared/createAi-mXofcZid.mjs";import{AI_DEFAULT_EMBEDDING_MODEL_ENV as a,AI_DEFAULT_MODEL_ENV as E,AI_GATEWAY_ACCOUNT_ID_ENV as o,AI_GATEWAY_ID_ENV as r,AI_GATEWAY_METADATA_MAX_KEYS as T,AI_GATEWAY_TAGS_ENV as I,AI_GATEWAY_TOKEN_ENV as l,buildAiGatewayMetadataFields as m,readAiGatewayEnvTags as s,resolveAiGateway as D}from"./packem_shared/AI_DEFAULT_EMBEDDING_MODEL_ENV-B58f6yo6.mjs";import{DEFAULT_MODEL_PRICES as M,estimateModelCost as d,lookupModelPrice as N}from"./packem_shared/DEFAULT_MODEL_PRICES-CIJLs1Ah.mjs";import{embed as x,embedMany as O,generateObject as c,generateText as f,hasToolCall as p,jsonSchema as L,streamObject as V,streamText as W,tool as Y}from"ai";import{createWorkersAI as n}from"workers-ai-provider";export{a as AI_DEFAULT_EMBEDDING_MODEL_ENV,E as AI_DEFAULT_MODEL_ENV,o as AI_GATEWAY_ACCOUNT_ID_ENV,r as AI_GATEWAY_ID_ENV,T as AI_GATEWAY_METADATA_MAX_KEYS,I as AI_GATEWAY_TAGS_ENV,l as AI_GATEWAY_TOKEN_ENV,M as DEFAULT_MODEL_PRICES,m as buildAiGatewayMetadataFields,_ as createAi,n as createWorkersAI,x as embed,O as embedMany,d as estimateModelCost,c as generateObject,f as generateText,p as hasToolCall,L as jsonSchema,N as lookupModelPrice,s as readAiGatewayEnvTags,D as resolveAiGateway,V as streamObject,W as streamText,Y as tool};

package/dist/packem_shared/AI_DEFAULT_EMBEDDING_MODEL_ENV-B58f6yo6.mjs

@@ -0,0 +1 @@
+const a=(t,o)=>{if(t===void 0)return;const e=t[o];return typeof e=="string"&&e.length>0?e:void 0};let A=!1,d=!1;const f=5,l=t=>{if(t===void 0)return;const o={};typeof t.functionPath=="string"&&t.functionPath.length>0&&(o.functionPath=t.functionPath),typeof t.traceId=="string"&&t.traceId.length>0&&(o.traceId=t.traceId);const e={};for(const[n,s]of Object.entries(t.tags??{}))typeof s=="string"&&s.length>0&&n.length>0&&!Object.hasOwn(o,n)&&(e[n]=s);Object.assign(e,o);const r=Object.keys(e);if(r.length===0)return;if(r.length<=f)return e;const i=r.slice(-f);return Object.fromEntries(i.map(n=>[n,e[n]]))},g=t=>{const o=l(t);return o===void 0?void 0:JSON.stringify(o)},I="LUNORA_AI_DEFAULT_MODEL",y="LUNORA_AI_DEFAULT_EMBEDDING_MODEL",E="LUNORA_AI_GATEWAY_ACCOUNT_ID",h="LUNORA_AI_GATEWAY_ID",_="LUNORA_AI_GATEWAY_TOKEN",u="LUNORA_AI_GATEWAY_TAGS",O=t=>{const o=a(t,u);if(o!==void 0)try{const e=JSON.parse(o);if(typeof e!="object"||e===null||Array.isArray(e))throw new TypeError("expected a JSON object");const r={};for(const[i,n]of Object.entries(e))typeof n=="string"&&n.length>0&&(r[i]=n);return Object.keys(r).length>0?r:void 0}catch{d||(d=!0,console.warn(`[lunora:ai] ${u} is not a flat JSON object of strings — AI Gateway tags from it are ignored.`));return}},T=(t,o,e="byo-provider")=>{const r=a(t,E),i=a(t,h);if(r===void 0||i===void 0)return;const n=a(t,_),s={};n!==void 0&&(s["cf-aig-authorization"]=`Bearer ${n}`,e==="workers-ai-binding"&&!A&&(A=!0,console.warn(`[lunora:ai] ${_} is set, but the Workers AI binding cannot send a gateway auth token — Cloudflare's native gateway option has no authorization field. The token is ignored on this path; use a bring-your-own AI SDK provider (which sends cf-aig-authorization), or make the AI Gateway unauthenticated for Workers AI.`)));const c=g(o);return c!==void 0&&(s["cf-aig-metadata"]=c),{accountId:r,baseURL:`https://gateway.ai.cloudflare.com/v1/${r}/${i}`,gatewayId:i,headers:s}};export{y as AI_DEFAULT_EMBEDDING_MODEL_ENV,I as AI_DEFAULT_MODEL_ENV,E as AI_GATEWAY_ACCOUNT_ID_ENV,h as AI_GATEWAY_ID_ENV,f as AI_GATEWAY_METADATA_MAX_KEYS,u as AI_GATEWAY_TAGS_ENV,_ as AI_GATEWAY_TOKEN_ENV,l as buildAiGatewayMetadataFields,O as readAiGatewayEnvTags,a as readEnv,T as resolveAiGateway};

package/dist/packem_shared/DEFAULT_MODEL_PRICES-CIJLs1Ah.mjs

@@ -0,0 +1 @@
+const p=/^(.*)-\d{4}-\d{2}-\d{2}$/u,a={"@cf/baai/bge-base-en-v1.5":{input:.067},"@cf/baai/bge-large-en-v1.5":{input:.204},"@cf/baai/bge-m3":{input:.012},"@cf/baai/bge-small-en-v1.5":{input:.02},"@cf/meta/llama-3.1-8b-instruct":{input:.28,output:.83},"@cf/meta/llama-3.3-70b-instruct-fp8-fast":{input:.29,output:2.25},"text-embedding-3-large":{input:.13},"text-embedding-3-small":{input:.02},"gpt-4o":{input:2.5,output:10},"gpt-4o-mini":{input:.15,output:.6},"gpt-5":{input:1.25,output:10}},c=e=>{const t=e.trim(),i=t.indexOf("/@"),n=i===-1?t:t.slice(i+1),u=n.lastIndexOf("/");return(u!==-1&&!n.startsWith("@")?[n,n.slice(u+1)]:[n]).flatMap(o=>{const r=p.exec(o)?.[1];return r===void 0?[o]:[o,r]})},b=(e,t=a)=>{for(const i of c(e))if(Object.hasOwn(t,i))return t[i]},d=(e,t,i)=>{if(e===void 0||e.length===0)return;const n=b(e,i);if(n===void 0)return;const u=Number.isFinite(t.inputTokens)?Math.max(0,t.inputTokens):0,s=Number.isFinite(t.outputTokens)?Math.max(0,t.outputTokens):0;if(u<=0&&s<=0)return;const o=(u*n.input+s*(n.output??0))/1e6;return Number.isFinite(o)?o:void 0};export{a as DEFAULT_MODEL_PRICES,d as estimateModelCost,b as lookupModelPrice};

package/dist/packem_shared/{createAi-CHSAfN98.mjs → createAi-mXofcZid.mjs}

@@ -1 +1 @@
-import{LunoraError as l}from"@lunora/errors";import{createWorkersAI as M}from"workers-ai-provider";import{readEnv as g,AI_DEFAULT_MODEL_ENV as w,AI_DEFAULT_EMBEDDING_MODEL_ENV as E,readAiGatewayEnvTags as y,buildAiGatewayMetadataFields as N,resolveAiGateway as h}from"./AI_DEFAULT_EMBEDDING_MODEL_ENV-BEAzcoBw.mjs";const D=(d,o,n)=>{const v=o===void 0?void 0:y(o),a=v===void 0?n:{...n,tags:{...v,...n?.tags}},t=N(a);if(d!==void 0)return t!==void 0&&d.metadata===void 0?{...d,metadata:t}:d;if(o===void 0)return;const i=h(o,a,"workers-ai-binding");if(i!==void 0)return t===void 0?{id:i.gatewayId}:{id:i.gatewayId,metadata:t}},R=d=>{const{binding:o,defaultEmbeddingModel:n,defaultModel:v,env:a,gateway:t,metadata:i,provider:c}=d;if(!c&&!o)throw new l("INTERNAL","@lunora/ai: createAi requires a `binding` (env.AI) or a pre-built `provider`");const f=D(t,a,i),m=v??g(a,w),b=n??g(a,E),s=c??M({binding:o,gateway:f}),A=e=>{if(e===void 0){if(!m)throw new l("INTERNAL",`@lunora/ai: no model supplied and no default configured — pass a model id, or set ${w} in the Worker env (wrangler \`vars\` / \`.dev.vars\`)`);return s(m)}return typeof e=="string"?s(e):e},p=e=>{const r=s.textEmbeddingModel;if(typeof r!="function")throw new l("INTERNAL","@lunora/ai: the Workers AI provider does not expose `textEmbeddingModel`; pass an AI SDK EmbeddingModel (e.g. from @ai-sdk/openai) to embed()");return r.call(s,e)};return{embeddingModel:e=>{if(typeof e=="object")return e;const r=e??b;if(!r)throw new l("INTERNAL",`@lunora/ai: no embedding model supplied and no default configured — pass an embedding model id or an AI SDK EmbeddingModel, or set ${E} in the Worker env (wrangler \`vars\` / \`.dev.vars\`)`);return p(r)},model:A,run:async(e,r,u)=>{if(!o)throw new l("INTERNAL","@lunora/ai: ai.run requires the `binding` (env.AI) — it is unavailable when only a custom `provider` was supplied");const I=f!==void 0&&u?.gateway===void 0?{...u,gateway:f}:u;return o.run(e,r,I)},workersai:s}};export{R as default};
+import{LunoraError as l}from"@lunora/errors";import{createWorkersAI as M}from"workers-ai-provider";import{readEnv as g,AI_DEFAULT_MODEL_ENV as w,AI_DEFAULT_EMBEDDING_MODEL_ENV as E,readAiGatewayEnvTags as y,buildAiGatewayMetadataFields as N,resolveAiGateway as h}from"./AI_DEFAULT_EMBEDDING_MODEL_ENV-B58f6yo6.mjs";const D=(d,o,n)=>{const v=o===void 0?void 0:y(o),a=v===void 0?n:{...n,tags:{...v,...n?.tags}},t=N(a);if(d!==void 0)return t!==void 0&&d.metadata===void 0?{...d,metadata:t}:d;if(o===void 0)return;const i=h(o,a,"workers-ai-binding");if(i!==void 0)return t===void 0?{id:i.gatewayId}:{id:i.gatewayId,metadata:t}},R=d=>{const{binding:o,defaultEmbeddingModel:n,defaultModel:v,env:a,gateway:t,metadata:i,provider:c}=d;if(!c&&!o)throw new l("INTERNAL","@lunora/ai: createAi requires a `binding` (env.AI) or a pre-built `provider`");const f=D(t,a,i),m=v??g(a,w),b=n??g(a,E),s=c??M({binding:o,gateway:f}),A=e=>{if(e===void 0){if(!m)throw new l("INTERNAL",`@lunora/ai: no model supplied and no default configured — pass a model id, or set ${w} in the Worker env (wrangler \`vars\` / \`.dev.vars\`)`);return s(m)}return typeof e=="string"?s(e):e},p=e=>{const r=s.textEmbeddingModel;if(typeof r!="function")throw new l("INTERNAL","@lunora/ai: the Workers AI provider does not expose `textEmbeddingModel`; pass an AI SDK EmbeddingModel (e.g. from @ai-sdk/openai) to embed()");return r.call(s,e)};return{embeddingModel:e=>{if(typeof e=="object")return e;const r=e??b;if(!r)throw new l("INTERNAL",`@lunora/ai: no embedding model supplied and no default configured — pass an embedding model id or an AI SDK EmbeddingModel, or set ${E} in the Worker env (wrangler \`vars\` / \`.dev.vars\`)`);return p(r)},model:A,run:async(e,r,u)=>{if(!o)throw new l("INTERNAL","@lunora/ai: ai.run requires the `binding` (env.AI) — it is unavailable when only a custom `provider` was supplied");const I=f!==void 0&&u?.gateway===void 0?{...u,gateway:f}:u;return o.run(e,r,I)},workersai:s}};export{R as default};

package/dist/packem_shared/defineRag-Cs7m4xJJ.mjs

@@ -0,0 +1,8 @@
+import{LunoraError as b,isLunoraError as _e}from"@lunora/errors";import{tool as Te,jsonSchema as Ae,embedMany as Me,embed as Ne}from"ai";import{s as De}from"./stable-key-B_BlboiY.mjs";import{estimateModelCost as Re}from"./DEFAULT_MODEL_PRICES-CIJLs1Ah.mjs";import Ce from"./fixedWindowChunks-C461ahRE.mjs";import{c as Be,I as Oe}from"./concurrent-C6nqBv41.mjs";import{contentHash as $e}from"./contentHash-BIn6ECP8.mjs";import{hybridRank as Ke}from"./hybridRank-B4skyCLx.mjs";import{VECTORIZE_CAPABILITIES as ae,vectorizeStore as Ue}from"./VECTORIZE_CAPABILITIES-CUQDoxis.mjs";const Qe=1e3,Ve=200,Fe=5,Le=4,ie=ae.maxMetadataBytes===!1?Number.POSITIVE_INFINITY:ae.maxMetadataBytes,je=2*1024,le="__ragChunk",me="__ragSource",C="__ragText",P="__ragHash",Y="__ragChunks",$="__ragImportance",he="__ragModel",ze=new Set([le,Y,P,$,he,me,C]),Pe=(e,d,h,w)=>{if(w===!1)return;const v=new TextEncoder().encode(JSON.stringify(e)).length;if(v<=w)return;const K=(typeof e[C]=="string"?new TextEncoder().encode(e[C]).length:0)*2>v?"lower `chunkSize` (it counts characters, not bytes — multibyte text costs up to 3 bytes each), or supply `textStore` to move chunk text out of metadata entirely":"attach less per-source `metadata`";throw new b("BAD_REQUEST",`@lunora/ai/rag: chunk ${String(d)} of "${h}" carries ${String(v)} bytes of metadata, over the store's ${String(w)}-byte per-vector ceiling — ${K}`)},Ye=(e,d,h)=>{if(h===!1)return;const w=new TextEncoder().encode(e).length;if(!(w<=h))throw new b("BAD_REQUEST",`@lunora/ai/rag: chunk id "${e}" for source "${d}" is ${String(w)} bytes, over the store's ${String(h)}-byte per-vector id ceiling — shorten the source id (hash long keys before indexing them) or shorten the \`namespace\`, which is prefixed onto every chunk id`)},He=/^[\w.-]{1,40}$/,pe=e=>e===void 0?"":`${encodeURIComponent(e)}#`,R=(e,d,h)=>`${pe(e)}${d}#${String(h)}`,de=(e,d)=>{const h=pe(d),w=h!==""&&e.startsWith(h)?e.slice(h.length):e,v=w.lastIndexOf("#"),A=v===-1?Number.NaN:Number(w.slice(v+1));return v===-1||!Number.isInteger(A)||A<0?{chunkIndex:0,sourceId:w}:{chunkIndex:A,sourceId:w.slice(0,v)}},qe=async e=>$e(new TextEncoder().encode(e)),We=e=>{try{return De([e.text,e.metadata,e.importance])}catch{return}},Ze=(e,d)=>{const h=[],w=[];for(const v of e)v.score>=d?h.push(v):w.push(v.id);return{kept:h,rejectedIds:w}},Xe=e=>e!==void 0&&Object.keys(e).length>0,j=e=>{if(!e)return;const d=Object.entries(e).filter(([h])=>!ze.has(h));return d.length>0?Object.fromEntries(d):void 0},ce=new Set,Je=e=>{ce.has(e)||(ce.add(e),console.warn(`[@lunora/ai/rag] index "${e}" is used without a namespace — in a multi-tenant/sharded
+app this shares one tenant's chunks (text included) with every other tenant, since
+Vectorize indexes are account-global. Pass \`namespace\` (the shard/tenant key) on both
+index() and retrieve(). Single-tenant apps suppress this via { allowSharedNamespace: true }.`))},Ge=e=>e.map(d=>`[source:${d.sourceId}#${String(d.chunkIndex)}]
+${d.text}`).join(`
+
+`),ue=(e,d)=>{if(typeof e=="object")return e;if(d===void 0)throw new b("INTERNAL","@lunora/ai/rag: `embeddingModel` is a Workers AI model id (or omitted) but the bound context has no `ai` (env.AI). Pass an AI SDK EmbeddingModel object to embed without Workers AI, or bind a context whose `ctx.ai` is wired.");return d.embeddingModel(e)},z=e=>{const d=e.modelId;return typeof d=="string"&&d.length>0?d:void 0},et=e=>{if(!(typeof e!="object"||e===null)){for(const d of Object.values(e))if(typeof d=="object"&&d!==null){const{cost:h}=d;if(typeof h=="number"&&Number.isFinite(h))return h}}},tt=e=>{if(e===void 0)throw new b("INTERNAL","@lunora/ai/rag: the bound context has no `vectors` (env.VECTORIZE) and no `store` is configured — bind a context whose `ctx.vectors` is wired, or configure `store` (e.g. `sqliteVectorStore`) to back this index without Vectorize.");return e},mt=e=>{if(typeof e.index!="string"||e.index.length===0)throw new b("BAD_REQUEST","@lunora/ai/rag: `index` must be a non-empty Vectorize index name");const d=e.chunkSize??Qe,h=e.chunkOverlap??Ve;if(!Number.isInteger(d)||d<1)throw new b("BAD_REQUEST","@lunora/ai/rag: `chunkSize` must be a positive integer");if(!Number.isInteger(h)||h<0||h>=d)throw new b("BAD_REQUEST","@lunora/ai/rag: `chunkOverlap` must be a non-negative integer smaller than `chunkSize`");const w=ie-je;if(!e.chunk&&!e.textStore&&!e.store&&d>w)throw new b("BAD_REQUEST",`@lunora/ai/rag: \`chunkSize\` of ${String(d)} leaves no room under Vectorize's ${String(ie)}-byte metadata limit, which also carries this chunk's bookkeeping and any \`metadata\` you attach — keep it under ${String(w)}, or supply \`textStore\` to move chunk text out of metadata entirely`);const v=e.topK??Fe;if(!Number.isInteger(v)||v<1)throw new b("BAD_REQUEST","@lunora/ai/rag: `topK` must be a positive integer");if(e.maxEmbeddingDimensions!==void 0&&e.maxEmbeddingDimensions!==!1&&(!Number.isInteger(e.maxEmbeddingDimensions)||e.maxEmbeddingDimensions<1))throw new b("BAD_REQUEST","@lunora/ai/rag: `maxEmbeddingDimensions` must be a positive integer, or `false` to disable the check");if(e.embeddingModelVersion!==void 0&&!He.test(e.embeddingModelVersion))throw new b("BAD_REQUEST",'@lunora/ai/rag: `embeddingModelVersion` must match /^[A-Za-z0-9._-]{1,40}$/ (a short, stable tag like "bge-v1.5")');if(e.candidates!==void 0&&(!Number.isInteger(e.candidates)||e.candidates<1))throw new b("BAD_REQUEST","@lunora/ai/rag: `candidates` must be a positive integer");if(e.cacheEmbeddings!==void 0&&(!Number.isInteger(e.cacheEmbeddings)||e.cacheEmbeddings<0))throw new b("BAD_REQUEST","@lunora/ai/rag: `cacheEmbeddings` must be a non-negative integer");const A=e.cacheEmbeddings??0,K=e.rerank,fe=e.chunk??(p=>Ce(p,d,h)),{textStore:k}=e,B=e.embeddingModelVersion,Q=p=>B===void 0?p:p===void 0?B:`${B}::${p}`;return p=>{const E=e.store?e.store(p):Ue(tt(p.vectors),e.index),H=k?E.capabilities.maxTopK:E.capabilities.maxTopKWithMetadata,U=e.maxEmbeddingDimensions??E.capabilities.maxDimensions;let O;const q=typeof p.trace=="function"?p.trace:void 0;let W=U===!1;const Z=(t,r)=>{if(W||(W=!0,U===!1||t<=U))return;const s=z(r);throw new b("BAD_REQUEST",`@lunora/ai/rag: embedding model${s===void 0?"":` "${s}"`} produces ${String(t)}-dimension vectors, over the ${String(U)}-dimension ceiling of index "${e.index}" — either truncate them with the provider's \`dimensions\` option (Matryoshka models such as text-embedding-3-large support this), or set \`maxEmbeddingDimensions: false\` if this index is not Vectorize-backed`)},N=new Map,ge=(t,r)=>{if(A!==0)for(N.set(t,r);N.size>A;){const s=N.keys().next();if(s.done===!0)break;N.delete(s.value)}},X=async t=>{const r=N.get(t);if(r!==void 0)return r;O??=ue(e.embeddingModel,p.ai);const s=O,a=async n=>{const{embedding:m,providerMetadata:y,usage:l}=await Ne({model:s,value:t});if(Z(m.length,s),n!==void 0){const c=l.tokens;typeof c=="number"&&Number.isFinite(c)&&n.setAttribute("gen_ai.usage.input_tokens",c);const u=et(y),f=u??Re(z(s),{inputTokens:typeof c=="number"?c:void 0});f!==void 0&&(n.setAttribute("gen_ai.usage.cost",f),n.setAttribute("lunora.usage.cost.source",u===void 0?"estimated":"provider"))}return ge(t,m),m};if(q===void 0)return a();const o=z(s),i=typeof p.conversationId=="string"&&p.conversationId.length>0?p.conversationId:void 0;return q("ai.embed",(n,m)=>a(m),{"gen_ai.operation.name":"embeddings",...o===void 0?{}:{"gen_ai.request.model":o},...i===void 0?{}:{"gen_ai.conversation.id":i}})},be=async(t,r,s)=>{if(!e.transformQuery||r?.transformQuery===!1)return[t];const a=typeof p.conversationId=="string"&&p.conversationId.length>0?p.conversationId:void 0,o=await e.transformQuery(t,{conversationId:a,namespace:s}),i=(typeof o=="string"?[o]:[...o]).map(n=>n.trim()).filter(n=>n.length>0);return i.length>0?i:[t]},we=async t=>{const r=new Map,s=[...new Set(t.filter(a=>!N.has(a)))];if(s.length<2)return r;O??=ue(e.embeddingModel,p.ai);try{const{embeddings:a}=await Me({model:O,values:s});if(a.length!==s.length)return r;const[o]=a;o!==void 0&&Z(o.length,O);for(const[i,n]of s.entries())r.set(n,a[i])}catch(a){if(_e(a))throw a}return r},V=t=>{if(t===void 0){if(e.requireNamespace)throw new b("BAD_REQUEST",`@lunora/ai/rag: index "${e.index}" requires a namespace (requireNamespace is set) — pass the tenant/shard key on index()/retrieve()/remove()`);e.allowSharedNamespace||Je(e.index)}},J=async(t,r)=>{const[s]=await E.getByIds([R(r,t,0)],r),a=s?.metadata?.[P],o=s?.metadata?.[Y];return{chunks:typeof o=="number"&&Number.isInteger(o)&&o>0?o:void 0,hash:typeof a=="string"?a:void 0}},G=async(t,r,s,a)=>{const o=Array.from({length:s-r},(i,n)=>R(a,t,r+n));o.length!==0&&(await E.deleteByIds(o,a),await k?.remove?.(o,{namespace:a}),await e.lexicalStore?.remove?.(o,{namespace:a}))},ve=async t=>{if(V(t.namespace),t.importance!==void 0&&(typeof t.importance!="number"||t.importance<0||t.importance>1))throw new b("BAD_REQUEST","@lunora/ai/rag: `importance` must be a number in [0, 1]");const r=Q(t.namespace),s=We(t),a=await qe(s??t.text),o=await J(t.id,r);if(t.reindex!==!0&&s!==void 0&&o.hash===a&&o.chunks!==void 0)return{chunks:o.chunks,ids:Array.from({length:o.chunks},(c,u)=>R(r,t.id,u)),unchanged:!0};const i=fe(t.text),n=i.map((c,u)=>R(r,t.id,u)),m=n.at(-1);if(m!==void 0&&Ye(m,t.id,E.capabilities.maxIdBytes),i.length===0&&t.allowEmptySources===!1)throw new b("BAD_REQUEST",`@lunora/ai/rag: source "${t.id}" produced zero chunks — set allowEmptySources: true to allow this`);if(i.length>0){const c=i.map((u,f)=>({chunkIndex:f,id:n[f],sourceId:t.id,text:u,...t.metadata===void 0?{}:{metadata:t.metadata}}));k&&await k.put(c,{namespace:r}),e.lexicalStore&&await e.lexicalStore.index(c,{namespace:r})}const y=await we(i),l=async c=>y.get(c)??await X(c);return await Be(i,Oe,async(c,u)=>{const f=n[u],x={...t.metadata,[le]:u,[me]:t.id};k||(x[C]=c),t.importance!==void 0&&(x[$]=t.importance),u===0&&(x[P]=a,x[Y]=i.length,B!==void 0&&(x[he]=B)),Pe(x,u,t.id,E.capabilities.maxMetadataBytes),await E.upsert({embed:l,id:f,input:c,metadata:x,namespace:r}),t.onChunk?.({chunkIndex:u,id:f,text:c,total:i.length})}),o.chunks!==void 0&&o.chunks>i.length&&await G(t.id,i.length,o.chunks,r),{chunks:i.length,ids:n,unchanged:!1}},ye=async t=>{V(t.namespace);const r=Q(t.namespace),a=(await J(t.id,r)).chunks??1;await G(t.id,0,a,r)},ee=async(t,r)=>{const s=new Map;if(t.length===0)return s;if(k){const o=await k.getMany(t,{namespace:r});for(const[i,n]of t.entries()){const m=o[i];typeof m=="string"&&s.set(n,m)}return s}const a=await E.getByIds(t,r);for(const o of a){const i=o.metadata?.[C];typeof i=="string"&&s.set(o.id,i)}return s},Ee=async(t,r,s)=>{const a=r?.chunkContext?.before??0,o=r?.chunkContext?.after??0;if(a===0&&o===0)return t;if(!Number.isInteger(a)||a<0||!Number.isInteger(o)||o<0)throw new b("BAD_REQUEST","@lunora/ai/rag: `chunkContext.before`/`chunkContext.after` must be non-negative integers");const i=new Map(t.map(l=>[l.id,l.text])),n=new Set;for(const l of t)for(let c=-a;c<=o;c+=1){const u=l.chunkIndex+c,f=R(s,l.sourceId,u);c!==0&&u>=0&&!i.has(f)&&n.add(f)}const m=await ee([...n],s),y=(l,c)=>{const u=R(s,l,c);return i.get(u)??m.get(u)};return t.map(l=>{const c=[];for(let u=-a;u<=o;u+=1){const f=u===0?l.text:y(l.sourceId,l.chunkIndex+u);f!==void 0&&c.push(f)}return{...l,text:c.join(`
+`)}})},xe=t=>{if(typeof t=="string"){const r=e.filters!==void 0&&Object.hasOwn(e.filters,t)?e.filters[t]:void 0;if(!r)throw new b("NOT_FOUND",`@lunora/ai/rag: unknown named filter "${t}" — must be one of the keys declared in RagConfig.filters`);return r.filter}return t},Ie=(t,r)=>t.matches.map(s=>{const a=s.metadata??{},o=de(s.id,r),i=a[C],n=a[$],m=typeof n=="number"&&n>=0&&n<=1?n:1;return{chunkIndex:o.chunkIndex,id:s.id,importance:m,metadata:j(a),score:s.score*m,sourceId:o.sourceId,text:typeof i=="string"?i:""}}),Se=async(t,r)=>{if(!k)return t;const s=t.map(n=>n.id),[a,o]=await Promise.all([ee(s,r),E.getByIds(s,r)]),i=new Map(o.map(n=>[n.id,n.metadata]));return t.flatMap(n=>{const m=a.get(n.id);if(m===void 0)return[];const y=i.get(n.id),l=y?.[$],c=typeof l=="number"&&l>=0&&l<=1?l:n.importance,f=(n.importance===0?0:n.score/n.importance)*c;return[{...n,importance:c,metadata:j(y)??n.metadata,score:f,text:m}]})},te=async(t,r,s)=>{const a=t.map(n=>n.id).filter(n=>!r.has(n)),o=a.length===0?[]:await E.getByIds(a,s),i=new Map(o.map(n=>[n.id,n.metadata]));return t.map(n=>{const m=de(n.id,s),y=i.get(n.id),l=y?.[$];return{chunkIndex:m.chunkIndex,id:n.id,importance:typeof l=="number"&&l>=0&&l<=1?l:1,metadata:j(y),score:n.score,sourceId:m.sourceId,text:n.text}})},ne=async(t,r)=>{V(r?.namespace);const s=Q(r?.namespace),a=xe(r?.filter),o=e.rlsFilter?await e.rlsFilter(p.auth):void 0,i=o?{...a,...o}:a,n=Math.min(r?.topK??v,H),m=await be(t,r,s),y=m[0],l=K!==void 0&&r?.rerank!==!1,u=l||e.lexicalStore!==void 0||e.graphStore!==void 0||m.length>1?Math.min(e.candidates??n*Le,H):n,f=r?.minScore,x=new Set,re=async g=>{const _=await E.query({embed:X,filter:i,input:g,namespace:s,returnMetadata:k?"indexed":"all",topK:u}),M=await Se(Ie(_,s),s);if(f===void 0)return M;const{kept:S,rejectedIds:T}=Ze(M,f);for(const ke of T)x.add(ke);return S},D=[{chunks:await re(y)}];for(const g of m.slice(1))D.push({chunks:await re(g)});if(e.lexicalStore){const g=await e.lexicalStore.search(y,{filter:i,namespace:s,topK:e.lexicalTopK??u}),_=new Set(D.flatMap(S=>S.chunks).map(S=>S.id)),M=g.filter(S=>!x.has(S.id)||_.has(S.id));D.push({chunks:await te(M,_,s)})}const F=D.flatMap(g=>[...g.chunks]);if(e.graphStore&&F.length>0&&(e.graphStore.enforcesFilter||!Xe(i))){const g=[...new Set(F.map(T=>T.sourceId))],_=await e.graphStore.related(g,{filter:i,namespace:s,topK:e.graphTopK??u}),M=new Set(F.map(T=>T.id)),S=_.filter(T=>!x.has(T.id)||M.has(T.id));D.push({chunks:await te(S,M,s),weight:"proximity"})}const L=D.filter(g=>g.chunks.length>0);let I=L.length>1?[...Ke(L)]:[...L[0]?.chunks??[]];I.sort((g,_)=>_.score-g.score),l&&(I=[...await K(y,I)]),I=I.slice(0,n),I=[...await Ee(I,r,s)];const se=[],oe=new Set;for(const g of I)oe.has(g.sourceId)||(oe.add(g.sourceId),se.push({id:g.sourceId,metadata:g.metadata,weight:g.importance}));return r?.onRetrieve?.({matches:I.length,query:t}),{chunks:I,context:Ge(I),sources:se}};return{asTool:t=>Te({description:t?.description??`Search the "${e.index}" knowledge base for passages relevant to a natural-language query.`,execute:async({query:r})=>ne(r,{namespace:t?.namespace,topK:t?.topK}),inputSchema:Ae({properties:{query:{description:"The natural-language search query.",type:"string"}},required:["query"],type:"object"})}),index:ve,remove:ye,retrieve:ne}}};export{mt as default};

package/dist/packem_shared/{markdownChunker-Bcv56GEz.mjs → markdownChunker-iW8V4klY.mjs}

@@ -1,5 +1,5 @@
-import m from"./fixedWindowChunks-C461ahRE.mjs";const E=1e3,T=200,w=256,b=/(?<=[!.?])\s/u,z=/^(#{1,6})\s(.*)$/u,C=/^ {0,3}(?:`{3,}|~{3,})/u,f=(o,r)=>{const e=o?.size??E,t=o?.overlap??T;if(!Number.isInteger(e)||e<1)throw new RangeError(`${r}: \`size\` must be a positive integer`);if(!Number.isInteger(t)||t<0||t>=e)throw new RangeError(`${r}: \`overlap\` must be a non-negative integer smaller than \`size\``);return{overlap:t,size:e}},k=o=>o.split(b).map(r=>r.trim()).filter(r=>r.length>0),x=(o,r)=>o.length===0?0:o.reduce((e,t)=>e+t.length,0)+r.length*(o.length-1),N=(o,r)=>{if(r.overlap===0)return[];const e=[];for(let t=o.length-1;t>=0&&e.length<o.length-1&&(e.unshift(o[t]),!(r.measure(e)>=r.overlap));t-=1);return e},d=(o,r)=>{const{budget:e,measure:t,separator:u,splitOversized:s}=r,c=[];let n=[];const i=()=>{n.length>0&&c.push(n.join(u))};for(const l of o)if(l.trim().length!==0){if(t([l])>e){i(),n=[],c.push(...s(l));continue}if(n.length>0&&t([...n,l])>e)for(i(),n=N(n,r);n.length>0&&t([...n,l])>e;)n.shift();n.push(l)}return i(),c.filter(l=>l.trim().length>0)},A=o=>{const r=[];let e=[],t=[],u=[],s=!1;const c=()=>{t.some(n=>n.trim().length>0)&&r.push({body:t,trail:u}),t=[]};for(const n of o.split(`
-`)){if(C.test(n)){s=!s,t.push(n);continue}const i=s?void 0:z.exec(n)??void 0;if(i===void 0){t.push(n);continue}c();const l=i[1].length,a=i[2].trim();e=[...e.slice(0,l-1),`${"#".repeat(l)} ${a}`],u=e}return c(),r},p=o=>{const{overlap:r,size:e}=f(o,"sentenceChunker");return t=>{const u=t.trim();return u.length===0?[]:d(k(u),{budget:e,measure:s=>x(s," "),overlap:r,separator:" ",splitOversized:s=>m(s,e,r)})}},O=o=>{const{overlap:r,size:e}=f(o,"markdownChunker"),t=p({overlap:r,size:e}),u=(s,c)=>{const n=e-c.length;return n<Math.ceil(e/4)?t(s):p({overlap:Math.min(r,n-1),size:n})(s).map(i=>`${c}${i}`)};return s=>{if(s.trim().length===0)return[];const c=[];for(const n of A(s)){const i=n.body.join(`
+import f from"./fixedWindowChunks-C461ahRE.mjs";const E=1e3,T=200,w=256,b=/(?<=[!.?])\s/u,z=/^(#{1,6})\s(.*)$/u,C=/^ {0,3}(?:`{3,}|~{3,})/u,k=(o,r)=>{const e=o?.size??E,t=o?.overlap??T;if(!Number.isInteger(e)||e<1)throw new RangeError(`${r}: \`size\` must be a positive integer`);if(!Number.isInteger(t)||t<0||t>=e)throw new RangeError(`${r}: \`overlap\` must be a non-negative integer smaller than \`size\``);return{overlap:t,size:e}},d=o=>o.split(b).map(r=>r.trim()).filter(r=>r.length>0),x=(o,r)=>o.length===0?0:o.reduce((e,t)=>e+t.length,0)+r.length*(o.length-1),N=(o,r)=>{if(r.overlap===0)return[];const e=[];for(let t=o.length-1;t>=0&&e.length<o.length-1&&(e.unshift(o[t]),!(r.measure(e)>=r.overlap));t-=1);return e},v=(o,r)=>{const{budget:e,measure:t,separator:u,splitOversized:s}=r,c=[];let n=[];const i=()=>{n.length>0&&c.push(n.join(u))};for(const l of o)if(l.trim().length!==0){if(t([l])>e){i(),n=[],c.push(...s(l));continue}if(n.length>0&&t([...n,l])>e)for(i(),n=N(n,r);n.length>0&&t([...n,l])>e;)n.shift();n.push(l)}return i(),c.filter(l=>l.trim().length>0)},A=o=>{const r=[];let e=[],t=[],u=[],s=!1;const c=()=>{t.some(n=>n.trim().length>0)&&r.push({body:t,trail:u}),t=[]};for(const n of o.split(`
+`)){if(C.test(n)){s=!s,t.push(n);continue}const i=s?void 0:z.exec(n)??void 0;if(i===void 0){t.push(n);continue}c();const l=i[1].length,a=i[2].trim();e=[...e.filter(h=>h.indexOf(" ")<l),`${"#".repeat(l)} ${a}`],u=e}return c(),r},m=o=>{const{overlap:r,size:e}=k(o,"sentenceChunker");return t=>{const u=t.trim();return u.length===0?[]:v(d(u),{budget:e,measure:s=>x(s," "),overlap:r,separator:" ",splitOversized:s=>f(s,e,r)})}},M=o=>{const{overlap:r,size:e}=k(o,"markdownChunker"),t=m({overlap:r,size:e}),u=(s,c)=>{const n=e-c.length;return n<Math.ceil(e/4)?t(s):m({overlap:Math.min(r,n-1),size:n})(s).map(i=>`${c}${i}`)};return s=>{if(s.trim().length===0)return[];const c=[];for(const n of A(s)){const i=n.body.join(`
 `).trim();i.length>0&&c.push(...u(i,n.trail.length>0?`${n.trail.join(" > ")}
 
-`:""))}return c}},S=o=>{const{countTokens:r}=o,e=o.maxTokens??w,t=o.overlapTokens??0;if(typeof r!="function")throw new TypeError("tokenChunker: `countTokens` must be a function — pass your model's tokenizer (e.g. js-tiktoken)");if(!Number.isInteger(e)||e<1)throw new RangeError("tokenChunker: `maxTokens` must be a positive integer");if(!Number.isInteger(t)||t<0||t>=e)throw new RangeError("tokenChunker: `overlapTokens` must be a non-negative integer smaller than `maxTokens`");const u=s=>{const c=[],n=[s];for(;n.length>0;){const i=n.pop(),l=r(i);if(l<=e||i.length<=1){c.push(i);continue}const a=Math.floor(i.length*e/Math.max(1,l)),v=Math.min(i.length-1,Math.max(1,a)),g=m(i,v,0);for(let h=g.length-1;h>=0;h-=1)n.push(g[h])}return c};return s=>{const c=s.trim();return c.length===0?[]:d(k(c),{budget:e,measure:n=>r(n.join(" ")),overlap:t,separator:" ",splitOversized:u})}};export{O as markdownChunker,p as sentenceChunker,S as tokenChunker};
+`:""))}return c}},S=o=>{const{countTokens:r}=o,e=o.maxTokens??w,t=o.overlapTokens??0;if(typeof r!="function")throw new TypeError("tokenChunker: `countTokens` must be a function — pass your model's tokenizer (e.g. js-tiktoken)");if(!Number.isInteger(e)||e<1)throw new RangeError("tokenChunker: `maxTokens` must be a positive integer");if(!Number.isInteger(t)||t<0||t>=e)throw new RangeError("tokenChunker: `overlapTokens` must be a non-negative integer smaller than `maxTokens`");const u=s=>{const c=[],n=[s];for(;n.length>0;){const i=n.pop(),l=r(i);if(l<=e||i.length<=1){c.push(i);continue}const a=Math.floor(i.length*e/Math.max(1,l)),h=Math.min(i.length-1,Math.max(1,a)),p=f(i,h,0);for(let g=p.length-1;g>=0;g-=1)n.push(p[g])}return c};return s=>{const c=s.trim();return c.length===0?[]:v(d(c),{budget:e,measure:n=>r(n.join(" ")),overlap:t,separator:" ",splitOversized:u})}};export{M as markdownChunker,m as sentenceChunker,S as tokenChunker};

package/dist/packem_shared/ragSyncTriggers-hgMcA4f2.mjs

@@ -0,0 +1 @@
+import{s as l}from"./stable-key-B_BlboiY.mjs";const v=c=>{const m=c.delayMs??0,y=(d,a)=>c.id===void 0?a:c.id(d),e=d=>d===void 0?void 0:c.text(d),u=d=>c.metadata?.(d),p=d=>c.namespace?.(d),o=(d,a)=>{const t=p(d);return{id:y(d,a),...t===void 0?{}:{namespace:t}}},f=(d,a,t)=>{const s=u(d);return{...o(d,a),...s===void 0?{}:{metadata:s},text:t}},g=(d,a)=>{try{return l(u(d))===l(u(a))}catch{return!1}},r=async(d,a)=>{await d.scheduler.runAfter(m,c.action,a)};return{afterDelete:async(d,a)=>{const t=a.previous??a.doc;await r(d,{deleted:!0,...t===void 0?{id:a.id}:o(t,a.id)})},afterInsert:async(d,a)=>{const t=e(a.doc);t===void 0||a.doc===void 0||await r(d,f(a.doc,a.id,t))},afterUpdate:async(d,a)=>{if(a.doc===void 0)return;const t=e(a.doc),s=e(a.previous),i=o(a.doc,a.id),n=a.previous===void 0?i:o(a.previous,a.id),w=n.id!==i.id||n.namespace!==i.namespace,h=c.metadata!==void 0&&a.previous!==void 0&&!g(a.doc,a.previous);if(w)await r(d,{deleted:!0,...n});else if(t===s&&!h)return;await(t===void 0?r(d,{deleted:!0,...i}):r(d,f(a.doc,a.id,t)))}}};export{v as ragSyncTriggers};

package/dist/packem_shared/stable-key-B_BlboiY.mjs

@@ -0,0 +1 @@
+const a=/["\\\u0000-\u001F\uD800-\uDFFF]/,f=t=>a.test(t)?JSON.stringify(t):`"${t}"`,y=t=>{if(t===void 0)return"null";if(typeof t=="bigint")throw new TypeError("stableStringify: cannot use a bigint in a stable JSON cache key — pass it as a string, or use stableWireKey");if(typeof t=="number"){if(Number.isNaN(t))return"nan";if(t===1/0)return"inf";if(t===-1/0)return"-inf";if(Object.is(t,-0))return"-0"}if(typeof t=="string")return f(t);if(t===null||typeof t!="object")return JSON.stringify(t);if(Array.isArray(t)){let r="[";for(let n=0;n<t.length;n++)n>0&&(r+=","),r+=y(t[n]);return r+"]"}const i=Object.getPrototypeOf(t);if(i!==null&&i!==Object.prototype){const r=t.constructor?.name??"value";throw new TypeError(`stableStringify: cannot use a ${r} in a stable JSON cache key — only plain objects, arrays, and JSON primitives are supported (wire-typed values key via stableWireKey)`)}const o=t,c=Object.keys(o).sort();let e="{",s=!0;for(const r of c){const n=o[r];n!==void 0&&(s?s=!1:e+=",",e+=f(r),e+=":",e+=y(n))}return e+"}"};export{y as s};

package/dist/rag/index.d.mts

@@ -1217,24 +1217,6 @@ interface SqliteVectorStoreOptions {
   table?: string;
 }
 declare const sqliteVectorStore: (options: SqliteVectorStoreOptions) => RagVectorStore;
-/**
- * Keep a RAG index in step with a table, so the table IS the index.
- *
- * `rag.index(...)` is a manual call, which means an app that edits a document
- * has to remember to re-index it — and a forgotten call is invisible: retrieval
- * keeps answering, just from stale text. The fix is to hang the re-index off the
- * write itself.
- *
- * Embedding is network I/O, so it cannot run inside the mutation that wrote the
- * row. The bridge is the two seams that already exist: a table `.triggers()`
- * handler runs in the write path and can `ctx.scheduler.runAfter(...)`, so the
- * trigger records the intent and an internal ACTION does the embedding a moment
- * later. That is what {@link ragSyncTriggers} wires up.
- *
- * Re-indexing unchanged text is already cheap — `rag.index` short-circuits on a
- * content hash — but this skips scheduling entirely when an update didn't touch
- * the indexed text, so an unrelated column edit costs nothing at all.
- */
 /**
  * A dispatchable function reference — the `internal.docs.reindex` you pass as
  * `action`. Typed structurally (rather than `unknown`) so passing the wrong
@@ -1260,10 +1242,14 @@ interface RagSyncEvent {
 }
 /** What the scheduled action receives — one document to re-index, or one to drop. */
 interface RagSyncArgs extends Record<string, unknown> {
-  /** `true` when the source row was deleted: call `rag.remove({ id })`. */
+  /** `true` when the source's chunks must go: call `rag.remove({ id, namespace })`. */
   deleted?: boolean;
   /** The source id — the same `id` you pass to `rag.index`/`rag.remove`. */
   id: string;
+  /** The {@link RagSyncOptions.metadata} projection, to pass to `rag.index`. Absent on a delete or when there is none. */
+  metadata?: Record<string, unknown>;
+  /** The {@link RagSyncOptions.namespace} projection, to pass to `rag.index`/`rag.remove`. Absent when there is none. */
+  namespace?: string;
   /** The text to embed. Absent on a delete. */
   text?: string;
 }
@@ -1282,6 +1268,19 @@ interface RagSyncOptions<Document extends Record<string, unknown> = Record<strin
   delayMs?: number;
   /** The source id to index under. Defaults to the row's own id. */
   id?: (document: Document) => string;
+  /**
+   * Metadata copied onto every chunk (`rag.index({ metadata })`) — the fields a
+   * `rlsFilter` or `metadataFilter` scopes retrieval by, such as a tenant or
+   * owner id. Compared on update like the text: a row that moves tenant with
+   * unchanged text is re-indexed, so the old tenant stops retrieving it.
+   */
+  metadata?: (document: Document) => Record<string, unknown> | undefined;
+  /**
+   * The Vectorize namespace to index under (`rag.index({ namespace })`), e.g.
+   * the tenant key. When it changes on update, the chunks in the old namespace
+   * are removed before the row is indexed into the new one.
+   */
+  namespace?: (document: Document) => string | undefined;
   /** The text to embed. Return `undefined` to skip the row (a draft, an empty body). */
   text: (document: Document) => string | undefined;
 }
@@ -1303,17 +1302,37 @@ type RagSyncHandler = (context: RagSyncTriggerContext, event: RagSyncEvent) => P
  * });
  * ```
  *
- * The action on the other end is three lines:
+ * The action on the other end forwards every field it receives:
  *
  * ```ts
- * export const reindex = internalAction.input({ deleted: v.optional(v.boolean()), id: v.string(), text: v.optional(v.string()) }).action(
- *     async ({ args, ctx }) => {
+ * export const reindex = internalAction
+ *     .input({
+ *         deleted: v.optional(v.boolean()),
+ *         id: v.string(),
+ *         metadata: v.optional(v.record(v.string(), v.any())),
+ *         namespace: v.optional(v.string()),
+ *         text: v.optional(v.string()),
+ *     })
+ *     .action(async ({ args, ctx }) => {
  *         const rag = docsRag(ctx);
+ *         const { id, metadata, namespace, text } = args;
  *
- *         await (args.deleted === true || args.text === undefined ? rag.remove({ id: args.id }) : rag.index({ id: args.id, text: args.text }));
- *     },
- * );
+ *         await (args.deleted === true || text === undefined ? rag.remove({ id, namespace }) : rag.index({ id, metadata, namespace, text }));
+ *     });
  * ```
+ *
+ * A multi-tenant table also passes `metadata` and/or `namespace` (for example
+ * `metadata: (doc) => ({ orgId: doc.orgId })`). Without them, a row that moves
+ * tenant keeps its old scope in the index and stays retrievable by the tenant it
+ * left. An action that drops `namespace` removes from (or indexes into) the
+ * wrong namespace, and one whose `input` omits `metadata` / `namespace` rejects
+ * the scheduled args and fails every job.
+ *
+ * Re-scoping is not atomic: the triggers only SCHEDULE the work, so after a row
+ * changes namespace its chunks stay retrievable in the old one until the
+ * scheduled delete runs (and until the re-index runs, it is missing from the new
+ * one). The same holds for a metadata change. Lower `delayMs` shortens that
+ * window; it does not close it.
  */
 declare const ragSyncTriggers: <Document extends Record<string, unknown> = Record<string, unknown>>(options: RagSyncOptions<Document>) => {
   afterDelete: RagSyncHandler;

package/dist/rag/index.d.ts

@@ -1217,24 +1217,6 @@ interface SqliteVectorStoreOptions {
   table?: string;
 }
 declare const sqliteVectorStore: (options: SqliteVectorStoreOptions) => RagVectorStore;
-/**
- * Keep a RAG index in step with a table, so the table IS the index.
- *
- * `rag.index(...)` is a manual call, which means an app that edits a document
- * has to remember to re-index it — and a forgotten call is invisible: retrieval
- * keeps answering, just from stale text. The fix is to hang the re-index off the
- * write itself.
- *
- * Embedding is network I/O, so it cannot run inside the mutation that wrote the
- * row. The bridge is the two seams that already exist: a table `.triggers()`
- * handler runs in the write path and can `ctx.scheduler.runAfter(...)`, so the
- * trigger records the intent and an internal ACTION does the embedding a moment
- * later. That is what {@link ragSyncTriggers} wires up.
- *
- * Re-indexing unchanged text is already cheap — `rag.index` short-circuits on a
- * content hash — but this skips scheduling entirely when an update didn't touch
- * the indexed text, so an unrelated column edit costs nothing at all.
- */
 /**
  * A dispatchable function reference — the `internal.docs.reindex` you pass as
  * `action`. Typed structurally (rather than `unknown`) so passing the wrong
@@ -1260,10 +1242,14 @@ interface RagSyncEvent {
 }
 /** What the scheduled action receives — one document to re-index, or one to drop. */
 interface RagSyncArgs extends Record<string, unknown> {
-  /** `true` when the source row was deleted: call `rag.remove({ id })`. */
+  /** `true` when the source's chunks must go: call `rag.remove({ id, namespace })`. */
   deleted?: boolean;
   /** The source id — the same `id` you pass to `rag.index`/`rag.remove`. */
   id: string;
+  /** The {@link RagSyncOptions.metadata} projection, to pass to `rag.index`. Absent on a delete or when there is none. */
+  metadata?: Record<string, unknown>;
+  /** The {@link RagSyncOptions.namespace} projection, to pass to `rag.index`/`rag.remove`. Absent when there is none. */
+  namespace?: string;
   /** The text to embed. Absent on a delete. */
   text?: string;
 }
@@ -1282,6 +1268,19 @@ interface RagSyncOptions<Document extends Record<string, unknown> = Record<strin
   delayMs?: number;
   /** The source id to index under. Defaults to the row's own id. */
   id?: (document: Document) => string;
+  /**
+   * Metadata copied onto every chunk (`rag.index({ metadata })`) — the fields a
+   * `rlsFilter` or `metadataFilter` scopes retrieval by, such as a tenant or
+   * owner id. Compared on update like the text: a row that moves tenant with
+   * unchanged text is re-indexed, so the old tenant stops retrieving it.
+   */
+  metadata?: (document: Document) => Record<string, unknown> | undefined;
+  /**
+   * The Vectorize namespace to index under (`rag.index({ namespace })`), e.g.
+   * the tenant key. When it changes on update, the chunks in the old namespace
+   * are removed before the row is indexed into the new one.
+   */
+  namespace?: (document: Document) => string | undefined;
   /** The text to embed. Return `undefined` to skip the row (a draft, an empty body). */
   text: (document: Document) => string | undefined;
 }
@@ -1303,17 +1302,37 @@ type RagSyncHandler = (context: RagSyncTriggerContext, event: RagSyncEvent) => P
  * });
  * ```
  *
- * The action on the other end is three lines:
+ * The action on the other end forwards every field it receives:
  *
  * ```ts
- * export const reindex = internalAction.input({ deleted: v.optional(v.boolean()), id: v.string(), text: v.optional(v.string()) }).action(
- *     async ({ args, ctx }) => {
+ * export const reindex = internalAction
+ *     .input({
+ *         deleted: v.optional(v.boolean()),
+ *         id: v.string(),
+ *         metadata: v.optional(v.record(v.string(), v.any())),
+ *         namespace: v.optional(v.string()),
+ *         text: v.optional(v.string()),
+ *     })
+ *     .action(async ({ args, ctx }) => {
  *         const rag = docsRag(ctx);
+ *         const { id, metadata, namespace, text } = args;
  *
- *         await (args.deleted === true || args.text === undefined ? rag.remove({ id: args.id }) : rag.index({ id: args.id, text: args.text }));
- *     },
- * );
+ *         await (args.deleted === true || text === undefined ? rag.remove({ id, namespace }) : rag.index({ id, metadata, namespace, text }));
+ *     });
  * ```
+ *
+ * A multi-tenant table also passes `metadata` and/or `namespace` (for example
+ * `metadata: (doc) => ({ orgId: doc.orgId })`). Without them, a row that moves
+ * tenant keeps its old scope in the index and stays retrievable by the tenant it
+ * left. An action that drops `namespace` removes from (or indexes into) the
+ * wrong namespace, and one whose `input` omits `metadata` / `namespace` rejects
+ * the scheduled args and fails every job.
+ *
+ * Re-scoping is not atomic: the triggers only SCHEDULE the work, so after a row
+ * changes namespace its chunks stay retrievable in the old one until the
+ * scheduled delete runs (and until the re-index runs, it is missing from the new
+ * one). The same holds for a metadata change. Lower `delayMs` shortens that
+ * window; it does not close it.
  */
 declare const ragSyncTriggers: <Document extends Record<string, unknown> = Record<string, unknown>>(options: RagSyncOptions<Document>) => {
   afterDelete: RagSyncHandler;

package/dist/rag/index.mjs

@@ -1 +1 @@
-import{default as o}from"../packem_shared/fixedWindowChunks-C461ahRE.mjs";import{markdownChunker as a,sentenceChunker as f,tokenChunker as n}from"../packem_shared/markdownChunker-Bcv56GEz.mjs";import{default as x}from"../packem_shared/defineRag-DK8Ifpn_.mjs";import{contentHash as p,guessMimeTypeFromExtension as i}from"../packem_shared/contentHash-BIn6ECP8.mjs";import{hybridRank as d}from"../packem_shared/hybridRank-B4skyCLx.mjs";import{default as k}from"../packem_shared/bm25LexicalStore-DMUzAL0O.mjs";import{default as h}from"../packem_shared/matchesMetadataFilter-BbIOyA5g.mjs";import{batchReranker as g,scoreReranker as C}from"../packem_shared/batchReranker-Bc38FBLH.mjs";import{defineRagSource as E}from"../packem_shared/defineRagSource-Q3f3niU8.mjs";import{sqlLexicalStore as T}from"../packem_shared/sqlLexicalStore-4C_cIwef.mjs";import{sqliteVectorStore as y}from"../packem_shared/sqliteVectorStore-D32l9lP0.mjs";import{ragSyncTriggers as q}from"../packem_shared/ragSyncTriggers-DPqzBNFw.mjs";import{VECTORIZE_CAPABILITIES as A,vectorizeStore as F}from"../packem_shared/VECTORIZE_CAPABILITIES-CUQDoxis.mjs";export{A as VECTORIZE_CAPABILITIES,g as batchReranker,k as bm25LexicalStore,p as contentHash,x as defineRag,E as defineRagSource,o as fixedWindowChunks,i as guessMimeTypeFromExtension,d as hybridRank,a as markdownChunker,h as matchesMetadataFilter,q as ragSyncTriggers,C as scoreReranker,f as sentenceChunker,T as sqlLexicalStore,y as sqliteVectorStore,n as tokenChunker,F as vectorizeStore};
+import{default as o}from"../packem_shared/fixedWindowChunks-C461ahRE.mjs";import{markdownChunker as a,sentenceChunker as f,tokenChunker as n}from"../packem_shared/markdownChunker-iW8V4klY.mjs";import{default as x}from"../packem_shared/defineRag-Cs7m4xJJ.mjs";import{contentHash as p,guessMimeTypeFromExtension as i}from"../packem_shared/contentHash-BIn6ECP8.mjs";import{hybridRank as d}from"../packem_shared/hybridRank-B4skyCLx.mjs";import{default as k}from"../packem_shared/bm25LexicalStore-DMUzAL0O.mjs";import{default as h}from"../packem_shared/matchesMetadataFilter-BbIOyA5g.mjs";import{batchReranker as g,scoreReranker as C}from"../packem_shared/batchReranker-Bc38FBLH.mjs";import{defineRagSource as E}from"../packem_shared/defineRagSource-Q3f3niU8.mjs";import{sqlLexicalStore as T}from"../packem_shared/sqlLexicalStore-4C_cIwef.mjs";import{sqliteVectorStore as y}from"../packem_shared/sqliteVectorStore-D32l9lP0.mjs";import{ragSyncTriggers as q}from"../packem_shared/ragSyncTriggers-hgMcA4f2.mjs";import{VECTORIZE_CAPABILITIES as A,vectorizeStore as F}from"../packem_shared/VECTORIZE_CAPABILITIES-CUQDoxis.mjs";export{A as VECTORIZE_CAPABILITIES,g as batchReranker,k as bm25LexicalStore,p as contentHash,x as defineRag,E as defineRagSource,o as fixedWindowChunks,i as guessMimeTypeFromExtension,d as hybridRank,a as markdownChunker,h as matchesMetadataFilter,q as ragSyncTriggers,C as scoreReranker,f as sentenceChunker,T as sqlLexicalStore,y as sqliteVectorStore,n as tokenChunker,F as vectorizeStore};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lunora/ai",
-  "version": "1.0.0-alpha.95",
+  "version": "1.0.0-alpha.97",
   "description": "Workers AI helper for Lunora: provider-agnostic AI SDK access from functions, Workers AI by default",
   "keywords": [
     "ai",
@@ -53,7 +53,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@lunora/errors": "1.0.0-alpha.40",
+    "@lunora/errors": "1.0.0-alpha.42",
     "ai": "7.0.93",
     "workers-ai-provider": "4.0.0"
   },

package/dist/packem_shared/AI_DEFAULT_EMBEDDING_MODEL_ENV-BEAzcoBw.mjs

@@ -1 +0,0 @@
-const a=(t,e)=>{if(t===void 0)return;const o=t[e];return typeof o=="string"&&o.length>0?o:void 0};let A=!1,d=!1;const f=5,l=t=>{if(t===void 0)return;const e={};for(const[n,r]of Object.entries(t.tags??{}))typeof r=="string"&&r.length>0&&n.length>0&&(e[n]=r);typeof t.functionPath=="string"&&t.functionPath.length>0&&(e.functionPath=t.functionPath),typeof t.traceId=="string"&&t.traceId.length>0&&(e.traceId=t.traceId);const o=Object.keys(e);if(o.length===0)return;if(o.length<=f)return e;const i=o.slice(-f);return Object.fromEntries(i.map(n=>[n,e[n]]))},E=t=>{const e=l(t);return e===void 0?void 0:JSON.stringify(e)},I="LUNORA_AI_DEFAULT_MODEL",y="LUNORA_AI_DEFAULT_EMBEDDING_MODEL",g="LUNORA_AI_GATEWAY_ACCOUNT_ID",h="LUNORA_AI_GATEWAY_ID",_="LUNORA_AI_GATEWAY_TOKEN",u="LUNORA_AI_GATEWAY_TAGS",T=t=>{const e=a(t,u);if(e!==void 0)try{const o=JSON.parse(e);if(typeof o!="object"||o===null||Array.isArray(o))throw new TypeError("expected a JSON object");const i={};for(const[n,r]of Object.entries(o))typeof r=="string"&&r.length>0&&(i[n]=r);return Object.keys(i).length>0?i:void 0}catch{d||(d=!0,console.warn(`[lunora:ai] ${u} is not a flat JSON object of strings — AI Gateway tags from it are ignored.`));return}},v=(t,e,o="byo-provider")=>{const i=a(t,g),n=a(t,h);if(i===void 0||n===void 0)return;const r=a(t,_),s={};r!==void 0&&(s["cf-aig-authorization"]=`Bearer ${r}`,o==="workers-ai-binding"&&!A&&(A=!0,console.warn(`[lunora:ai] ${_} is set, but the Workers AI binding cannot send a gateway auth token — Cloudflare's native gateway option has no authorization field. The token is ignored on this path; use a bring-your-own AI SDK provider (which sends cf-aig-authorization), or make the AI Gateway unauthenticated for Workers AI.`)));const c=E(e);return c!==void 0&&(s["cf-aig-metadata"]=c),{accountId:i,baseURL:`https://gateway.ai.cloudflare.com/v1/${i}/${n}`,gatewayId:n,headers:s}};export{y as AI_DEFAULT_EMBEDDING_MODEL_ENV,I as AI_DEFAULT_MODEL_ENV,g as AI_GATEWAY_ACCOUNT_ID_ENV,h as AI_GATEWAY_ID_ENV,f as AI_GATEWAY_METADATA_MAX_KEYS,u as AI_GATEWAY_TAGS_ENV,_ as AI_GATEWAY_TOKEN_ENV,l as buildAiGatewayMetadataFields,T as readAiGatewayEnvTags,a as readEnv,v as resolveAiGateway};

package/dist/packem_shared/DEFAULT_MODEL_PRICES-Q8uxdiuV.mjs

@@ -1 +0,0 @@
-const p=/^(.*)-\d{4}-\d{2}-\d{2}$/u,a={"@cf/baai/bge-base-en-v1.5":{input:.067},"@cf/baai/bge-large-en-v1.5":{input:.204},"@cf/baai/bge-m3":{input:.012},"@cf/baai/bge-small-en-v1.5":{input:.02},"@cf/meta/llama-3.1-8b-instruct":{input:.28,output:.83},"@cf/meta/llama-3.3-70b-instruct-fp8-fast":{input:.29,output:2.25},"text-embedding-3-large":{input:.13},"text-embedding-3-small":{input:.02},"gpt-4o":{input:2.5,output:10},"gpt-4o-mini":{input:.15,output:.6},"gpt-5":{input:1.25,output:10}},r=n=>{const t=n.trim(),i=t.lastIndexOf("/");return(i!==-1&&!t.startsWith("@")?[t,t.slice(i+1)]:[t]).flatMap(e=>{const o=p.exec(e)?.[1];return o===void 0?[e]:[e,o]})},c=(n,t=a)=>{for(const i of r(n))if(Object.hasOwn(t,i))return t[i]},b=(n,t,i)=>{if(n===void 0||n.length===0)return;const u=c(n,i);if(u===void 0)return;const e=Number.isFinite(t.inputTokens)?Math.max(0,t.inputTokens):0,o=Number.isFinite(t.outputTokens)?Math.max(0,t.outputTokens):0;if(e<=0&&o<=0)return;const s=(e*u.input+o*(u.output??0))/1e6;return Number.isFinite(s)?s:void 0};export{a as DEFAULT_MODEL_PRICES,b as estimateModelCost,c as lookupModelPrice};

package/dist/packem_shared/defineRag-DK8Ifpn_.mjs

@@ -1,8 +0,0 @@
-import{LunoraError as w,isLunoraError as Ne}from"@lunora/errors";import{tool as Ae,jsonSchema as Me,embedMany as Oe,embed as De}from"ai";import{estimateModelCost as Re}from"./DEFAULT_MODEL_PRICES-Q8uxdiuV.mjs";import Ce from"./fixedWindowChunks-C461ahRE.mjs";import{c as $e,I as Be}from"./concurrent-C6nqBv41.mjs";import{contentHash as Ke}from"./contentHash-BIn6ECP8.mjs";import{hybridRank as Ue}from"./hybridRank-B4skyCLx.mjs";import{VECTORIZE_CAPABILITIES as ce,vectorizeStore as Fe}from"./VECTORIZE_CAPABILITIES-CUQDoxis.mjs";const Qe=/["\\\u0000-\u001F\uD800-\uDFFF]/,de=e=>Qe.test(e)?JSON.stringify(e):`"${e}"`,Y=e=>{if(e===void 0)return"null";if(typeof e=="bigint")throw new TypeError("stableStringify: cannot use a bigint in a stable JSON cache key — pass it as a string, or use stableWireKey");if(typeof e=="number"){if(Number.isNaN(e))return"nan";if(e===1/0)return"inf";if(e===-1/0)return"-inf";if(Object.is(e,-0))return"-0"}if(typeof e=="string")return de(e);if(e===null||typeof e!="object")return JSON.stringify(e);if(Array.isArray(e)){let v="[";for(let T=0;T<e.length;T++)T>0&&(v+=","),v+=Y(e[T]);return v+"]"}const c=Object.getPrototypeOf(e);if(c!==null&&c!==Object.prototype){const v=e.constructor?.name??"value";throw new TypeError(`stableStringify: cannot use a ${v} in a stable JSON cache key — only plain objects, arrays, and JSON primitives are supported (wire-typed values key via stableWireKey)`)}const m=e,p=Object.keys(m).sort();let f="{",k=!0;for(const v of p){const T=m[v];T!==void 0&&(k?k=!1:f+=",",f+=de(v),f+=":",f+=Y(T))}return f+"}"},Ve=1e3,je=200,Le=5,ze=4,ue=ce.maxMetadataBytes===!1?Number.POSITIVE_INFINITY:ce.maxMetadataBytes,Pe=2*1024,fe="__ragChunk",pe="__ragSource",$="__ragText",q="__ragHash",H="__ragChunks",U="__ragImportance",ge="__ragModel",Ye=new Set([fe,H,q,U,ge,pe,$]),qe=(e,c,m,p)=>{if(p===!1)return;const f=new TextEncoder().encode(JSON.stringify(e)).length;if(f<=p)return;const v=(typeof e[$]=="string"?new TextEncoder().encode(e[$]).length:0)*2>f?"lower `chunkSize` (it counts characters, not bytes — multibyte text costs up to 3 bytes each), or supply `textStore` to move chunk text out of metadata entirely":"attach less per-source `metadata`";throw new w("BAD_REQUEST",`@lunora/ai/rag: chunk ${String(c)} of "${m}" carries ${String(f)} bytes of metadata, over the store's ${String(p)}-byte per-vector ceiling — ${v}`)},He=(e,c,m)=>{if(m===!1)return;const p=new TextEncoder().encode(e).length;if(!(p<=m))throw new w("BAD_REQUEST",`@lunora/ai/rag: chunk id "${e}" for source "${c}" is ${String(p)} bytes, over the store's ${String(m)}-byte per-vector id ceiling — shorten the source id (hash long keys before indexing them) or shorten the \`namespace\`, which is prefixed onto every chunk id`)},We=/^[\w.-]{1,40}$/,be=e=>e===void 0?"":`${encodeURIComponent(e)}#`,C=(e,c,m)=>`${be(e)}${c}#${String(m)}`,le=(e,c)=>{const m=be(c),p=m!==""&&e.startsWith(m)?e.slice(m.length):e,f=p.lastIndexOf("#"),k=f===-1?Number.NaN:Number(p.slice(f+1));return f===-1||!Number.isInteger(k)||k<0?{chunkIndex:0,sourceId:p}:{chunkIndex:k,sourceId:p.slice(0,f)}},Je=async e=>Ke(new TextEncoder().encode(e)),Ze=e=>{try{return Y([e.text,e.metadata,e.importance])}catch{return}},Xe=(e,c)=>{const m=[],p=[];for(const f of e)f.score>=c?m.push(f):p.push(f.id);return{kept:m,rejectedIds:p}},Ge=e=>e!==void 0&&Object.keys(e).length>0,z=e=>{if(!e)return;const c=Object.entries(e).filter(([m])=>!Ye.has(m));return c.length>0?Object.fromEntries(c):void 0},me=new Set,et=e=>{me.has(e)||(me.add(e),console.warn(`[@lunora/ai/rag] index "${e}" is used without a namespace — in a multi-tenant/sharded
-app this shares one tenant's chunks (text included) with every other tenant, since
-Vectorize indexes are account-global. Pass \`namespace\` (the shard/tenant key) on both
-index() and retrieve(). Single-tenant apps suppress this via { allowSharedNamespace: true }.`))},tt=e=>e.map(c=>`[source:${c.sourceId}#${String(c.chunkIndex)}]
-${c.text}`).join(`
-
-`),he=(e,c)=>{if(typeof e=="object")return e;if(c===void 0)throw new w("INTERNAL","@lunora/ai/rag: `embeddingModel` is a Workers AI model id (or omitted) but the bound context has no `ai` (env.AI). Pass an AI SDK EmbeddingModel object to embed without Workers AI, or bind a context whose `ctx.ai` is wired.");return c.embeddingModel(e)},P=e=>{const c=e.modelId;return typeof c=="string"&&c.length>0?c:void 0},nt=e=>{if(!(typeof e!="object"||e===null)){for(const c of Object.values(e))if(typeof c=="object"&&c!==null){const{cost:m}=c;if(typeof m=="number"&&Number.isFinite(m))return m}}},rt=e=>{if(e===void 0)throw new w("INTERNAL","@lunora/ai/rag: the bound context has no `vectors` (env.VECTORIZE) and no `store` is configured — bind a context whose `ctx.vectors` is wired, or configure `store` (e.g. `sqliteVectorStore`) to back this index without Vectorize.");return e},ht=e=>{if(typeof e.index!="string"||e.index.length===0)throw new w("BAD_REQUEST","@lunora/ai/rag: `index` must be a non-empty Vectorize index name");const c=e.chunkSize??Ve,m=e.chunkOverlap??je;if(!Number.isInteger(c)||c<1)throw new w("BAD_REQUEST","@lunora/ai/rag: `chunkSize` must be a positive integer");if(!Number.isInteger(m)||m<0||m>=c)throw new w("BAD_REQUEST","@lunora/ai/rag: `chunkOverlap` must be a non-negative integer smaller than `chunkSize`");const p=ue-Pe;if(!e.chunk&&!e.textStore&&!e.store&&c>p)throw new w("BAD_REQUEST",`@lunora/ai/rag: \`chunkSize\` of ${String(c)} leaves no room under Vectorize's ${String(ue)}-byte metadata limit, which also carries this chunk's bookkeeping and any \`metadata\` you attach — keep it under ${String(p)}, or supply \`textStore\` to move chunk text out of metadata entirely`);const f=e.topK??Le;if(!Number.isInteger(f)||f<1)throw new w("BAD_REQUEST","@lunora/ai/rag: `topK` must be a positive integer");if(e.maxEmbeddingDimensions!==void 0&&e.maxEmbeddingDimensions!==!1&&(!Number.isInteger(e.maxEmbeddingDimensions)||e.maxEmbeddingDimensions<1))throw new w("BAD_REQUEST","@lunora/ai/rag: `maxEmbeddingDimensions` must be a positive integer, or `false` to disable the check");if(e.embeddingModelVersion!==void 0&&!We.test(e.embeddingModelVersion))throw new w("BAD_REQUEST",'@lunora/ai/rag: `embeddingModelVersion` must match /^[A-Za-z0-9._-]{1,40}$/ (a short, stable tag like "bge-v1.5")');if(e.candidates!==void 0&&(!Number.isInteger(e.candidates)||e.candidates<1))throw new w("BAD_REQUEST","@lunora/ai/rag: `candidates` must be a positive integer");if(e.cacheEmbeddings!==void 0&&(!Number.isInteger(e.cacheEmbeddings)||e.cacheEmbeddings<0))throw new w("BAD_REQUEST","@lunora/ai/rag: `cacheEmbeddings` must be a non-negative integer");const k=e.cacheEmbeddings??0,v=e.rerank,T=e.chunk??(g=>Ce(g,c,m)),{textStore:N}=e,B=e.embeddingModelVersion,Q=g=>B===void 0?g:g===void 0?B:`${B}::${g}`;return g=>{const x=e.store?e.store(g):Fe(rt(g.vectors),e.index),W=N?x.capabilities.maxTopK:x.capabilities.maxTopKWithMetadata,F=e.maxEmbeddingDimensions??x.capabilities.maxDimensions;let K;const J=typeof g.trace=="function"?g.trace:void 0;let Z=F===!1;const X=(t,r)=>{if(Z||(Z=!0,F===!1||t<=F))return;const s=P(r);throw new w("BAD_REQUEST",`@lunora/ai/rag: embedding model${s===void 0?"":` "${s}"`} produces ${String(t)}-dimension vectors, over the ${String(F)}-dimension ceiling of index "${e.index}" — either truncate them with the provider's \`dimensions\` option (Matryoshka models such as text-embedding-3-large support this), or set \`maxEmbeddingDimensions: false\` if this index is not Vectorize-backed`)},D=new Map,ye=(t,r)=>{if(k!==0)for(D.set(t,r);D.size>k;){const s=D.keys().next();if(s.done===!0)break;D.delete(s.value)}},G=async t=>{const r=D.get(t);if(r!==void 0)return r;K??=he(e.embeddingModel,g.ai);const s=K,a=async n=>{const{embedding:h,providerMetadata:E,usage:l}=await De({model:s,value:t});if(X(h.length,s),n!==void 0){const d=l.tokens;typeof d=="number"&&Number.isFinite(d)&&n.setAttribute("gen_ai.usage.input_tokens",d);const u=nt(E),b=u??Re(P(s),{inputTokens:typeof d=="number"?d:void 0});b!==void 0&&(n.setAttribute("gen_ai.usage.cost",b),n.setAttribute("lunora.usage.cost.source",u===void 0?"estimated":"provider"))}return ye(t,h),h};if(J===void 0)return a();const o=P(s),i=typeof g.conversationId=="string"&&g.conversationId.length>0?g.conversationId:void 0;return J("ai.embed",(n,h)=>a(h),{"gen_ai.operation.name":"embeddings",...o===void 0?{}:{"gen_ai.request.model":o},...i===void 0?{}:{"gen_ai.conversation.id":i}})},we=async(t,r,s)=>{if(!e.transformQuery||r?.transformQuery===!1)return[t];const a=typeof g.conversationId=="string"&&g.conversationId.length>0?g.conversationId:void 0,o=await e.transformQuery(t,{conversationId:a,namespace:s}),i=(typeof o=="string"?[o]:[...o]).map(n=>n.trim()).filter(n=>n.length>0);return i.length>0?i:[t]},Ee=async t=>{const r=new Map,s=[...new Set(t.filter(a=>!D.has(a)))];if(s.length<2)return r;K??=he(e.embeddingModel,g.ai);try{const{embeddings:a}=await Oe({model:K,values:s});if(a.length!==s.length)return r;const[o]=a;o!==void 0&&X(o.length,K);for(const[i,n]of s.entries())r.set(n,a[i])}catch(a){if(Ne(a))throw a}return r},V=t=>{if(t===void 0){if(e.requireNamespace)throw new w("BAD_REQUEST",`@lunora/ai/rag: index "${e.index}" requires a namespace (requireNamespace is set) — pass the tenant/shard key on index()/retrieve()/remove()`);e.allowSharedNamespace||et(e.index)}},ee=async(t,r)=>{const[s]=await x.getByIds([C(r,t,0)],r),a=s?.metadata?.[q],o=s?.metadata?.[H];return{chunks:typeof o=="number"&&Number.isInteger(o)&&o>0?o:void 0,hash:typeof a=="string"?a:void 0}},te=async(t,r,s,a)=>{const o=Array.from({length:s-r},(i,n)=>C(a,t,r+n));o.length!==0&&(await x.deleteByIds(o,a),await N?.remove?.(o,{namespace:a}),await e.lexicalStore?.remove?.(o,{namespace:a}))},ve=async t=>{if(V(t.namespace),t.importance!==void 0&&(typeof t.importance!="number"||t.importance<0||t.importance>1))throw new w("BAD_REQUEST","@lunora/ai/rag: `importance` must be a number in [0, 1]");const r=Q(t.namespace),s=Ze(t),a=await Je(s??t.text),o=await ee(t.id,r);if(t.reindex!==!0&&s!==void 0&&o.hash===a&&o.chunks!==void 0)return{chunks:o.chunks,ids:Array.from({length:o.chunks},(d,u)=>C(r,t.id,u)),unchanged:!0};const i=T(t.text),n=i.map((d,u)=>C(r,t.id,u)),h=n.at(-1);if(h!==void 0&&He(h,t.id,x.capabilities.maxIdBytes),i.length===0&&t.allowEmptySources===!1)throw new w("BAD_REQUEST",`@lunora/ai/rag: source "${t.id}" produced zero chunks — set allowEmptySources: true to allow this`);if(i.length>0){const d=i.map((u,b)=>({chunkIndex:b,id:n[b],sourceId:t.id,text:u,...t.metadata===void 0?{}:{metadata:t.metadata}}));N&&await N.put(d,{namespace:r}),e.lexicalStore&&await e.lexicalStore.index(d,{namespace:r})}const E=await Ee(i),l=async d=>E.get(d)??await G(d);return await $e(i,Be,async(d,u)=>{const b=n[u],S={...t.metadata,[fe]:u,[pe]:t.id};N||(S[$]=d),t.importance!==void 0&&(S[U]=t.importance),u===0&&(S[q]=a,S[H]=i.length,B!==void 0&&(S[ge]=B)),qe(S,u,t.id,x.capabilities.maxMetadataBytes),await x.upsert({embed:l,id:b,input:d,metadata:S,namespace:r}),t.onChunk?.({chunkIndex:u,id:b,text:d,total:i.length})}),o.chunks!==void 0&&o.chunks>i.length&&await te(t.id,i.length,o.chunks,r),{chunks:i.length,ids:n,unchanged:!1}},xe=async t=>{V(t.namespace);const r=Q(t.namespace),a=(await ee(t.id,r)).chunks??1;await te(t.id,0,a,r)},ne=async(t,r)=>{const s=new Map;if(t.length===0)return s;if(N){const o=await N.getMany(t,{namespace:r});for(const[i,n]of t.entries()){const h=o[i];typeof h=="string"&&s.set(n,h)}return s}const a=await x.getByIds(t,r);for(const o of a){const i=o.metadata?.[$];typeof i=="string"&&s.set(o.id,i)}return s},Se=async(t,r,s)=>{const a=r?.chunkContext?.before??0,o=r?.chunkContext?.after??0;if(a===0&&o===0)return t;if(!Number.isInteger(a)||a<0||!Number.isInteger(o)||o<0)throw new w("BAD_REQUEST","@lunora/ai/rag: `chunkContext.before`/`chunkContext.after` must be non-negative integers");const i=new Map(t.map(l=>[l.id,l.text])),n=new Set;for(const l of t)for(let d=-a;d<=o;d+=1){const u=l.chunkIndex+d,b=C(s,l.sourceId,u);d!==0&&u>=0&&!i.has(b)&&n.add(b)}const h=await ne([...n],s),E=(l,d)=>{const u=C(s,l,d);return i.get(u)??h.get(u)};return t.map(l=>{const d=[];for(let u=-a;u<=o;u+=1){const b=u===0?l.text:E(l.sourceId,l.chunkIndex+u);b!==void 0&&d.push(b)}return{...l,text:d.join(`
-`)}})},Ie=t=>{if(typeof t=="string"){const r=e.filters!==void 0&&Object.hasOwn(e.filters,t)?e.filters[t]:void 0;if(!r)throw new w("NOT_FOUND",`@lunora/ai/rag: unknown named filter "${t}" — must be one of the keys declared in RagConfig.filters`);return r.filter}return t},ke=(t,r)=>t.matches.map(s=>{const a=s.metadata??{},o=le(s.id,r),i=a[$],n=a[U],h=typeof n=="number"&&n>=0&&n<=1?n:1;return{chunkIndex:o.chunkIndex,id:s.id,importance:h,metadata:z(a),score:s.score*h,sourceId:o.sourceId,text:typeof i=="string"?i:""}}),_e=async(t,r)=>{if(!N)return t;const s=t.map(n=>n.id),[a,o]=await Promise.all([ne(s,r),x.getByIds(s,r)]),i=new Map(o.map(n=>[n.id,n.metadata]));return t.flatMap(n=>{const h=a.get(n.id);if(h===void 0)return[];const E=i.get(n.id),l=E?.[U],d=typeof l=="number"&&l>=0&&l<=1?l:n.importance,b=(n.importance===0?0:n.score/n.importance)*d;return[{...n,importance:d,metadata:z(E)??n.metadata,score:b,text:h}]})},re=async(t,r,s)=>{const a=t.map(n=>n.id).filter(n=>!r.has(n)),o=a.length===0?[]:await x.getByIds(a,s),i=new Map(o.map(n=>[n.id,n.metadata]));return t.map(n=>{const h=le(n.id,s),E=i.get(n.id),l=E?.[U];return{chunkIndex:h.chunkIndex,id:n.id,importance:typeof l=="number"&&l>=0&&l<=1?l:1,metadata:z(E),score:n.score,sourceId:h.sourceId,text:n.text}})},se=async(t,r)=>{V(r?.namespace);const s=Q(r?.namespace),a=Ie(r?.filter),o=e.rlsFilter?await e.rlsFilter(g.auth):void 0,i=o?{...a,...o}:a,n=Math.min(r?.topK??f,W),h=await we(t,r,s),E=h[0],l=v!==void 0&&r?.rerank!==!1,u=l||e.lexicalStore!==void 0||e.graphStore!==void 0||h.length>1?Math.min(e.candidates??n*ze,W):n,b=r?.minScore,S=new Set,oe=async y=>{const A=await x.query({embed:G,filter:i,input:y,namespace:s,returnMetadata:N?"indexed":"all",topK:u}),O=await _e(ke(A,s),s);if(b===void 0)return O;const{kept:_,rejectedIds:M}=Xe(O,b);for(const Te of M)S.add(Te);return _},R=[{chunks:await oe(E)}];for(const y of h.slice(1))R.push({chunks:await oe(y)});if(e.lexicalStore){const y=await e.lexicalStore.search(E,{filter:i,namespace:s,topK:e.lexicalTopK??u}),A=new Set(R.flatMap(_=>_.chunks).map(_=>_.id)),O=y.filter(_=>!S.has(_.id)||A.has(_.id));R.push({chunks:await re(O,A,s)})}const j=R.flatMap(y=>[...y.chunks]);if(e.graphStore&&j.length>0&&(e.graphStore.enforcesFilter||!Ge(i))){const y=[...new Set(j.map(M=>M.sourceId))],A=await e.graphStore.related(y,{filter:i,namespace:s,topK:e.graphTopK??u}),O=new Set(j.map(M=>M.id)),_=A.filter(M=>!S.has(M.id)||O.has(M.id));R.push({chunks:await re(_,O,s),weight:"proximity"})}const L=R.filter(y=>y.chunks.length>0);let I=L.length>1?[...Ue(L)]:[...L[0]?.chunks??[]];I.sort((y,A)=>A.score-y.score),l&&(I=[...await v(E,I)]),I=I.slice(0,n),I=[...await Se(I,r,s)];const ae=[],ie=new Set;for(const y of I)ie.has(y.sourceId)||(ie.add(y.sourceId),ae.push({id:y.sourceId,metadata:y.metadata,weight:y.importance}));return r?.onRetrieve?.({matches:I.length,query:t}),{chunks:I,context:tt(I),sources:ae}};return{asTool:t=>Ae({description:t?.description??`Search the "${e.index}" knowledge base for passages relevant to a natural-language query.`,execute:async({query:r})=>se(r,{namespace:t?.namespace,topK:t?.topK}),inputSchema:Me({properties:{query:{description:"The natural-language search query.",type:"string"}},required:["query"],type:"object"})}),index:ve,remove:xe,retrieve:se}}};export{ht as default};

package/dist/packem_shared/ragSyncTriggers-DPqzBNFw.mjs

@@ -1 +0,0 @@
-const l=o=>{const u=o.delayMs??0,s=(i,d)=>o.id===void 0?d:o.id(i),t=i=>i===void 0?void 0:o.text(i),c=async(i,d)=>{await i.scheduler.runAfter(u,o.action,d)};return{afterDelete:async(i,d)=>{const r=d.previous??d.doc;await c(i,{deleted:!0,id:r===void 0?d.id:s(r,d.id)})},afterInsert:async(i,d)=>{const r=t(d.doc);r===void 0||d.doc===void 0||await c(i,{id:s(d.doc,d.id),text:r})},afterUpdate:async(i,d)=>{if(d.doc===void 0)return;const r=t(d.doc),f=t(d.previous),a=s(d.doc,d.id),e=d.previous===void 0?a:s(d.previous,d.id);if(e!==a)await c(i,{deleted:!0,id:e});else if(r===f)return;await(r===void 0?c(i,{deleted:!0,id:a}):c(i,{id:a,text:r}))}}};export{l as ragSyncTriggers};