Better streaming

- Always stream the visible scenarios, if the modelProvider supports it
 - Never stream the invisible scenarios

Also actually runs our query tasks in a background worker, which we weren't quite doing before.

src/server/utils/generateNewCell.ts

@@ -1,12 +1,18 @@
 import { type Prisma } from "@prisma/client";
 import { prisma } from "../db";
-import { queueLLMRetrievalTask } from "./queueLLMRetrievalTask";
 import parseConstructFn from "./parseConstructFn";
 import { type JsonObject } from "type-fest";
 import hashPrompt from "./hashPrompt";
 import { omit } from "lodash-es";
+import { queueQueryModel } from "../tasks/queryModel.task";
+
+export const generateNewCell = async (
+  variantId: string,
+  scenarioId: string,
+  options?: { stream?: boolean },
+): Promise<void> => {
+  const stream = options?.stream ?? false;
 
-export const generateNewCell = async (variantId: string, scenarioId: string): Promise<void> => {
   const variant = await prisma.promptVariant.findUnique({
     where: {
       id: variantId,
@@ -98,6 +104,6 @@ export const generateNewCell = async (variantId: string, scenarioId: string): Pr
       }),
     );
   } else {
-    cell = await queueLLMRetrievalTask(cell.id);
+    await queueQueryModel(cell.id, stream);
   }
 };

src/server/utils/queueLLMRetrievalTask.ts

@@ -1,22 +0,0 @@
-import { prisma } from "../db";
-import { queryLLM } from "../tasks/queryLLM.task";
-
-export const queueLLMRetrievalTask = async (cellId: string) => {
-  const updatedCell = await prisma.scenarioVariantCell.update({
-    where: {
-      id: cellId,
-    },
-    data: {
-      retrievalStatus: "PENDING",
-      errorMessage: null,
-    },
-    include: {
-      modelOutput: true,
-    },
-  });
-
-  // @ts-expect-error we aren't passing the helpers but that's ok
-  void queryLLM.task.handler({ scenarioVariantCellId: cellId }, { logger: console });
-
-  return updatedCell;
-};