@wei840222/qmd 2026.9.27 → 2026.9.28

package/dist/cli/build-info.json

@@ -1,4 +1,4 @@
 {
-  "commit": "ffd0c56",
-  "builtAt": "2026-09-27T09:25:17.053Z"
+  "commit": "65f6a34",
+  "builtAt": "2026-09-27T15:15:31.166Z"
 }

package/dist/hybrid.d.ts

@@ -11,6 +11,8 @@ export declare class Hybrid implements LLM {
     get local(): LLM;
     get supportsExpand(): boolean;
     get supportsRerank(): boolean;
+    get rerankModelName(): string | undefined;
+    get generateModelName(): string | undefined;
     embed(text: string, options?: EmbedOptions): Promise<EmbeddingResult | null>;
     generate(prompt: string, options?: GenerateOptions): Promise<GenerateResult | null>;
     modelExists(model: string): Promise<ModelInfo>;

package/dist/hybrid.js

@@ -22,6 +22,21 @@ export class Hybrid {
     get supportsRerank() {
         return Boolean(this.remoteJev?.supportsRerank || this.remoteLLM?.supportsRerank);
     }
+    get rerankModelName() {
+        if (this.remoteJev?.supportsRerank) {
+            return this.remoteJev.rerankModelName;
+        }
+        if (this.remoteLLM?.supportsRerank && this.remoteLLM.rerankModelName) {
+            return this.remoteLLM.rerankModelName;
+        }
+        return this.localLLM.rerankModelName;
+    }
+    get generateModelName() {
+        if (this.remoteLLM?.supportsExpand && this.remoteLLM.generateModelName) {
+            return this.remoteLLM.generateModelName;
+        }
+        return this.localLLM.generateModelName;
+    }
     async embed(text, options) {
         return this.localLLM.embed(text, options);
     }

package/dist/llm.d.ts

@@ -114,6 +114,8 @@ export type GenerateOptions = {
  */
 export type RerankOptions = {
     model?: string;
+    timeZone?: string;
+    context?: string;
 };
 /**
  * Options for LLM sessions

package/dist/remote-jev.d.ts

@@ -1,6 +1,7 @@
 import { TypeSafeClient } from "@typesafe-ai/sdk";
 import type { RerankDocument, RerankOptions, RerankResult, SearchIntentGuidance } from "./llm.js";
 export declare const DEFAULT_JEV_TIMEOUT_MS = 30000;
+export declare const DEFAULT_JEV_RERANK_BATCH_SIZE = 40;
 export interface JevSystemOneClient {
     systemOne(request: any, options?: any): Promise<any>;
 }
@@ -9,6 +10,7 @@ export interface RemoteJevOptions {
     baseUrl?: string;
     model?: string;
     concurrency?: number;
+    batchSize?: number;
     timeoutMs?: number;
     client?: TypeSafeClient | JevSystemOneClient;
 }
@@ -24,10 +26,12 @@ export interface JevIntentClassification {
 export declare class RemoteJev {
     readonly model: string;
     readonly concurrency: number;
+    readonly batchSize: number;
     readonly client: TypeSafeClient | JevSystemOneClient;
     constructor(options?: RemoteJevOptions);
     get supportsRerank(): boolean;
     get supportsExpand(): boolean;
+    get rerankModelName(): string;
     resetCircuitBreaker(): void;
     classifyIntent(query: string, options?: {
         context?: string;

package/dist/remote-jev.js

@@ -1,6 +1,7 @@
 import { TypeSafeClient, choice, noul } from "@typesafe-ai/sdk";
 import { getFormattedLocalTime } from "./remote-llm.js";
 export const DEFAULT_JEV_TIMEOUT_MS = 30000;
+export const DEFAULT_JEV_RERANK_BATCH_SIZE = 40;
 export const JEV_STRATEGY_PLAYBOOK = {
     code_search: {
         label: "Code Search",
@@ -30,10 +31,12 @@ export const JEV_STRATEGY_PLAYBOOK = {
 export class RemoteJev {
     model;
     concurrency;
+    batchSize;
     client;
     constructor(options = {}) {
         this.model = options.model?.trim() || "jev-1.13";
         this.concurrency = options.concurrency ?? 10;
+        this.batchSize = options.batchSize ?? DEFAULT_JEV_RERANK_BATCH_SIZE;
         if (options.client) {
             this.client = options.client;
         }
@@ -54,6 +57,9 @@ export class RemoteJev {
     get supportsExpand() {
         return true;
     }
+    get rerankModelName() {
+        return `jev:${this.model}`;
+    }
     resetCircuitBreaker() {
         // Kept for interface compatibility; error handling is handled per-request in caller/fallback
     }
@@ -92,38 +98,63 @@ export class RemoteJev {
         if (documents.length === 0) {
             return { results: [], model: `jev:${this.model}` };
         }
-        const rerankQuestion = noul("Does this candidate document answer or address the search query?", {
+        const state = { query };
+        const timeZone = typeof options === "object" && options !== null ? options.timeZone : undefined;
+        state.current_time = getFormattedLocalTime(new Date(), timeZone);
+        if (typeof options === "object" && options !== null && options.context) {
+            state.context = options.context;
+        }
+        const criteria = {
             true: "The candidate directly addresses the query's specific question, requirement, or topic, satisfying any time or entity constraints.",
             false: "The candidate is only on a similar topic, outside the requested time window, or unrelated to the query's specific need.",
-        });
-        const results = await pMap(documents, async (doc, index) => {
-            const text = typeof doc === "string" ? doc : doc.text;
-            const file = typeof doc === "string" ? doc : doc.file;
-            const candidate = truncateCandidateText(text, 1500);
-            const state = { query, candidate };
-            if (typeof doc !== "string" && doc.title) {
-                state.title = doc.title;
-            }
-            if (typeof doc !== "string" && doc.file) {
-                state.file = doc.file;
+        };
+        const batches = [];
+        for (let offset = 0; offset < documents.length; offset += this.batchSize) {
+            batches.push({ docs: documents.slice(offset, offset + this.batchSize), offset });
+        }
+        const batchResults = await pMap(batches, async ({ docs: batchDocs, offset }) => {
+            const questions = {};
+            for (let i = 0; i < batchDocs.length; i++) {
+                const doc = batchDocs[i];
+                const text = typeof doc === "string" ? doc : doc.text;
+                const candidate = truncateCandidateText(text, 1500);
+                const instructions = {
+                    candidate,
+                    question: "Does `candidate` directly answer or address the search query in `query`?",
+                };
+                if (typeof doc !== "string" && doc.title) {
+                    instructions.title = doc.title;
+                }
+                if (typeof doc !== "string" && doc.file) {
+                    instructions.file = doc.file;
+                }
+                questions[`cand_${i}`] = noul(instructions, criteria);
             }
-            const timeZone = typeof options === "object" && options !== null ? options.timeZone : undefined;
-            state.current_time = getFormattedLocalTime(new Date(), timeZone);
             const response = await this.client.systemOne({
                 state,
                 model: this.model,
-                questions: { is_relevant: rerankQuestion },
+                questions,
             });
-            return {
-                file,
-                score: response.answers.is_relevant.noul,
-                index,
-            };
+            const results = [];
+            for (let i = 0; i < batchDocs.length; i++) {
+                const doc = batchDocs[i];
+                const file = typeof doc === "string" ? doc : doc.file;
+                const key = `cand_${i}`;
+                const answer = response?.answers?.[key];
+                const score = typeof answer?.noul === "number" ? answer.noul : 0;
+                results.push({
+                    file,
+                    score,
+                    index: offset + i,
+                });
+            }
+            return results;
         }, this.concurrency);
+        const flattened = batchResults.flat();
         // Sort by score descending
-        results.sort((a, b) => b.score - a.score);
+        flattened.sort((a, b) => b.score - a.score);
         return {
-            results,
+            results: flattened,
             model: `jev:${this.model}`,
         };
     }

package/dist/remote-llm.d.ts

@@ -33,6 +33,8 @@ export declare class RemoteLLM implements LLM {
     constructor(options: RemoteLLMOptions);
     get supportsExpand(): boolean;
     get supportsRerank(): boolean;
+    get rerankModelName(): string | undefined;
+    get generateModelName(): string | undefined;
     embed(_text: string, _options?: EmbedOptions): Promise<EmbeddingResult | null>;
     generate(_prompt: string, _options?: GenerateOptions): Promise<GenerateResult | null>;
     modelExists(_model: string): Promise<ModelInfo>;

package/dist/remote-llm.js

@@ -112,6 +112,12 @@ export class RemoteLLM {
     get supportsRerank() {
         return Boolean(this.rerankApiUrl && this.rerankApiModel && !this.rerankCircuitBroken);
     }
+    get rerankModelName() {
+        return this.rerankApiModel;
+    }
+    get generateModelName() {
+        return this.generateApiModel;
+    }
     async embed(_text, _options) {
         // Embedding is handled separately by EmbeddingProvider
         return null;

package/dist/store.js

@@ -2711,11 +2711,19 @@ export function createStore(dbPath, options = {}) {
         searchFTS: (query, limit, collectionName, filter) => searchFTS(db, query, limit, collectionName, filter),
         searchVec: (query, model, limit, collectionFilter, session, precomputedEmbedding, filter) => searchVec(db, query, model, limit, collectionFilter, session, precomputedEmbedding, store.embeddingProvider, store.authorizeRemoteRequest, store.llm, filter),
         // Query expansion & reranking
-        expandQuery: (query, model, expansionContext, options) => expandQuery(query, model ?? store.localLlm?.generateModelName ?? store.llm?.generateModelName ?? DEFAULT_QUERY_MODEL, db, expansionContext, store.llm, options),
-        invalidateExpansionCache: (query, expansionContext, options) => deleteExpansionCacheEntry(db, query, store.localLlm?.generateModelName ?? store.llm?.generateModelName ?? DEFAULT_QUERY_MODEL, expansionContext, options),
+        expandQuery: (query, model, expansionContext, options) => {
+            const activeGenerateModel = store.llm?.generateModelName ?? store.localLlm?.generateModelName ?? DEFAULT_QUERY_MODEL;
+            return expandQuery(query, model ?? activeGenerateModel, db, expansionContext, store.llm, options);
+        },
+        invalidateExpansionCache: (query, expansionContext, options) => {
+            const activeGenerateModel = store.llm?.generateModelName ?? store.localLlm?.generateModelName ?? DEFAULT_QUERY_MODEL;
+            return deleteExpansionCacheEntry(db, query, activeGenerateModel, expansionContext, options);
+        },
         rerank: (query, documents, model, rerankContext) => {
             const llm = getLlm(store);
-            return rerank(query, documents, model ?? store.localLlm?.rerankModelName ?? llm?.rerankModelName ?? DEFAULT_RERANK_MODEL, db, rerankContext, store.llm ?? llm);
+            const activeLlm = store.llm ?? llm;
+            const activeRerankModel = activeLlm?.rerankModelName ?? store.localLlm?.rerankModelName ?? llm?.rerankModelName ?? DEFAULT_RERANK_MODEL;
+            return rerank(query, documents, model ?? activeRerankModel, db, rerankContext, activeLlm);
         },
         // Document retrieval
         findDocument: (filename, options) => findDocument(db, filename, options),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wei840222/qmd",
-  "version": "2026.9.27",
+  "version": "2026.9.28",
   "packageManager": "pnpm@11.15.1",
   "description": "Query Markup Documents - On-device hybrid search for markdown files with BM25, vector search, and LLM reranking",
   "type": "module",