entity isolation and estimateTokenFix

packages/orchestration-service/src/chat/index.js

@@ -177,8 +177,11 @@ function fuseEpisodeResults(semanticEps, keywordEps, { semanticWeight, keywordWe
 }
 
 function estimateTokens(episode) {
-    return episode.token_count
-        ?? Math.ceil((episode.user_message.length + episode.ai_response.length) / 4);
+    //NOTE: episode.token_count is not used here.  It stores the
+    //full inference cost of that turn(prompt + injected memory + response),
+    //this inflate the episodes apparent size and starving the budget.
+    //Therefore, char/4 on the actual stored text is an honest selectedl
+    return Math.ceil((episode.user_message.length + episode.ai_response.length) / 4);
 }
 
 function buildScoredPool(fusedWithScores, recentEpisodes, entityBoostedIds, { entityWeight }) {

packages/orchestration-service/src/services/qdrant.js

@@ -32,10 +32,18 @@ async function searchEpisodes( vector, {limit = ORCHESTRATION.RECENT_EPISODE_LIM
 async function searchEntities(vector, { limit = ORCHESTRATION.ENTITIES_LIMIT, scoreThreshold = ORCHESTRATION.ENTITIES_THRESHOLD, projectId = undefined } = {}) {
     const body = { vector, limit, score_threshold: scoreThreshold, with_payload: true };
 
+    // non-project chats must also be filters to the "no project" pool
     if (projectId !== null && projectId !== undefined) {
         body.filter = {
             must: [{ key: 'projectId', match: { value: projectId } }]
         };
+    } else {
+        //entities from project sessions carry a projectId;
+        //without this branch, a non-project chat searches ALL entities and project knowledge leaks into the common pool.
+        //is_empty matches null AND missing payload keys, so pre-isolation-era entities are covered too
+        body.filter = {
+            must: [{ is_empty: {key: 'projectId'} }]
+        }
     }
     const res = await fetch(
         `${BASE_URL}/collections/${COLLECTIONS.ENTITIES}/points/search`,