@loreai/core 0.16.0 → 0.17.0

package/src/temporal.ts

@@ -1,6 +1,7 @@
 import { db, ensureProject } from "./db";
-import { ftsQuery, ftsQueryOr, EMPTY_QUERY } from "./search";
+import { ftsQuery, EMPTY_QUERY, runRelaxedSearch } from "./search";
 import { sanitizeSurrogates } from "./markdown";
+import * as embedding from "./embedding";
 import type { LoreMessage, LorePart } from "./types";
 import { isTextPart, isReasoningPart, isToolPart } from "./types";
 
@@ -93,6 +94,10 @@ export function store(input: {
         messageMetadata(input.info, input.parts),
         input.info.id,
       );
+    // Re-embed on content update (fire-and-forget)
+    if (embedding.isAvailable()) {
+      embedding.embedTemporalMessage(input.info.id, content);
+    }
     return;
   }
 
@@ -111,6 +116,11 @@ export function store(input: {
       input.info.time.created,
       messageMetadata(input.info, input.parts),
     );
+
+  // Embed new message for vector search (fire-and-forget)
+  if (embedding.isAvailable()) {
+    embedding.embedTemporalMessage(input.info.id, content);
+  }
 }
 
 export type TemporalMessage = {
@@ -156,7 +166,7 @@ export function markDistilled(ids: string[]) {
   const placeholders = ids.map(() => "?").join(",");
   db()
     .query(
-      `UPDATE temporal_messages SET distilled = 1 WHERE id IN (${placeholders})`,
+      `UPDATE temporal_messages SET distilled = 1, embedding = NULL WHERE id IN (${placeholders})`,
     )
     .run(...ids);
 }
@@ -194,8 +204,6 @@ export function search(input: {
 }): TemporalMessage[] {
   const pid = ensureProject(input.projectPath);
   const limit = input.limit ?? 20;
-  const q = ftsQuery(input.query);
-  if (q === EMPTY_QUERY) return [];
 
   const ftsSQL = input.sessionID
     ? `SELECT m.* FROM temporal_fts f
@@ -206,24 +214,14 @@ export function search(input: {
        CROSS JOIN temporal_messages m ON m.rowid = f.rowid
        WHERE f.content MATCH ? AND m.project_id = ?
        ORDER BY rank LIMIT ?`;
-  const params = input.sessionID
-    ? [q, pid, input.sessionID, limit]
-    : [q, pid, limit];
+
   try {
-    const results = db()
-      .query(ftsSQL)
-      .all(...params) as TemporalMessage[];
-    if (results.length) return results;
-
-    // AND returned nothing — try OR fallback for broader recall
-    const qOr = ftsQueryOr(input.query);
-    if (qOr === EMPTY_QUERY) return [];
-    const paramsOr = input.sessionID
-      ? [qOr, pid, input.sessionID, limit]
-      : [qOr, pid, limit];
-    return db()
-      .query(ftsSQL)
-      .all(...paramsOr) as TemporalMessage[];
+    return runRelaxedSearch(input.query, (matchExpr) => {
+      const params = input.sessionID
+        ? [matchExpr, pid, input.sessionID, limit]
+        : [matchExpr, pid, limit];
+      return db().query(ftsSQL).all(...params) as TemporalMessage[];
+    });
   } catch {
     // FTS5 still choked (edge case) — fall back to LIKE search
     return searchLike({
@@ -249,8 +247,6 @@ export function searchScored(input: {
 }): ScoredTemporalMessage[] {
   const pid = ensureProject(input.projectPath);
   const limit = input.limit ?? 20;
-  const q = ftsQuery(input.query);
-  if (q === EMPTY_QUERY) return [];
 
   const ftsSQL = input.sessionID
     ? `SELECT m.*, rank FROM temporal_fts f
@@ -261,20 +257,14 @@ export function searchScored(input: {
        CROSS JOIN temporal_messages m ON m.rowid = f.rowid
        WHERE f.content MATCH ? AND m.project_id = ?
        ORDER BY rank LIMIT ?`;
-  const params = input.sessionID
-    ? [q, pid, input.sessionID, limit]
-    : [q, pid, limit];
 
   try {
-    const results = db().query(ftsSQL).all(...params) as ScoredTemporalMessage[];
-    if (results.length) return results;
-
-    const qOr = ftsQueryOr(input.query);
-    if (qOr === EMPTY_QUERY) return [];
-    const paramsOr = input.sessionID
-      ? [qOr, pid, input.sessionID, limit]
-      : [qOr, pid, limit];
-    return db().query(ftsSQL).all(...paramsOr) as ScoredTemporalMessage[];
+    return runRelaxedSearch(input.query, (matchExpr) => {
+      const params = input.sessionID
+        ? [matchExpr, pid, input.sessionID, limit]
+        : [matchExpr, pid, limit];
+      return db().query(ftsSQL).all(...params) as ScoredTemporalMessage[];
+    });
   } catch {
     return [];
   }
@@ -348,6 +338,23 @@ export function undistilledCount(
   ).count;
 }
 
+/** Sum of estimated tokens across undistilled messages for a project/session. */
+export function undistilledTokens(
+  projectPath: string,
+  sessionID?: string,
+): number {
+  const pid = ensureProject(projectPath);
+  const query = sessionID
+    ? "SELECT COALESCE(SUM(tokens), 0) as total FROM temporal_messages WHERE project_id = ? AND session_id = ? AND distilled = 0"
+    : "SELECT COALESCE(SUM(tokens), 0) as total FROM temporal_messages WHERE project_id = ? AND distilled = 0";
+  const params = sessionID ? [pid, sessionID] : [pid];
+  return (
+    db()
+      .query(query)
+      .get(...params) as { total: number }
+  ).total;
+}
+
 export type PruneResult = {
   /** Rows deleted by the TTL pass (distilled=1 AND older than retention period). */
   ttlDeleted: number;

package/src/types.ts

@@ -245,6 +245,21 @@ export interface LLMClient {
        * auth through their own mechanisms.
        */
       sessionID?: string;
+      /**
+       * Maximum output tokens for this call. When absent, the adapter
+       * uses its built-in default (typically 8192).
+       *
+       * Worker call sites should set this explicitly based on expected
+       * output size to avoid wasting tokens on unnecessarily large
+       * output budgets.
+       *
+       * Adapter behavior:
+       * - Gateway: uses as `max_tokens` in the API request body
+       * - Pi: passes as `maxTokens` to `complete()`
+       * - OpenCode: cannot honor — SDK has no maxTokens on session.prompt();
+       *   the field is silently ignored
+       */
+      maxTokens?: number;
     },
   ): Promise<string | null>;
 }

package/src/worker-model.ts

@@ -1,15 +1,15 @@
 /**
  * Worker model resolution.
  *
- * Background workers (distillation, curation, query expansion) use the session
- * model by default. An explicit `workerModel` config override is supported for
- * cases where the user wants to pin background work to a specific model.
+ * Background workers (distillation, curation, query expansion) default to
+ * sonnet-4-6 when the session model is more expensive ($1.50+/M input).
+ * Sonnet-4-6 produces equivalent-quality distillations at lower cost.
+ * An explicit `workerModel` config override takes priority over this default.
  *
- * Previously this module contained dynamic worker model selection with
- * candidate discovery, two-phase validation (structural check + LLM judge),
- * and fingerprint-based staleness detection. That complexity was removed in
- * favor of always using the session model — A/B testing showed the quality
- * gap on complex conversations wasn't worth the infrastructure cost.
+ * Resolution order:
+ *   1. Explicit config override (`workerModel`)
+ *   2. Cost-aware default (sonnet-4 for expensive session models)
+ *   3. Session model fallback (same model as the conversation)
  */
 
 // ---------------------------------------------------------------------------
@@ -35,16 +35,21 @@ export type ModelInfo = {
 
 /**
  * Resolve the effective worker model for a given provider.
- * Priority: explicit config override > session model (fallback).
+ * Priority: explicit config override > cost-aware default > session model.
  */
 export function resolveWorkerModel(
   _providerID: string,
   configWorkerModel?: { providerID: string; modelID: string },
   configModel?: { providerID: string; modelID: string },
+  costAwareDefault?: { providerID: string; modelID: string },
 ): { providerID: string; modelID: string } | undefined {
   // Explicit override wins
   if (configWorkerModel) return configWorkerModel;
 
+  // Cost-aware default: cheaper model for background work when the session
+  // model is expensive. Caller determines when this applies based on pricing.
+  if (costAwareDefault) return costAwareDefault;
+
   // Fall back to the session model config (or undefined = host default)
   return configModel;
 }