cache eval outputs and add gpt4 eval

prisma/migrations/20230718005301_add_gpt4_eval_type/migration.sql

@@ -0,0 +1,39 @@
+/*
+  Warnings:
+
+  - You are about to drop the `EvaluationResult` table. If the table is not empty, all the data it contains will be lost.
+
+*/
+-- AlterEnum
+ALTER TYPE "EvalType" ADD VALUE 'GPT4_EVAL';
+
+-- DropForeignKey
+ALTER TABLE "EvaluationResult" DROP CONSTRAINT "EvaluationResult_evaluationId_fkey";
+
+-- DropForeignKey
+ALTER TABLE "EvaluationResult" DROP CONSTRAINT "EvaluationResult_promptVariantId_fkey";
+
+-- DropTable
+DROP TABLE "EvaluationResult";
+
+-- CreateTable
+CREATE TABLE "OutputEvaluation" (
+    "id" UUID NOT NULL,
+    "result" DOUBLE PRECISION NOT NULL,
+    "details" TEXT,
+    "modelOutputId" UUID NOT NULL,
+    "evaluationId" UUID NOT NULL,
+    "createdAt" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    "updatedAt" TIMESTAMP(3) NOT NULL,
+
+    CONSTRAINT "OutputEvaluation_pkey" PRIMARY KEY ("id")
+);
+
+-- CreateIndex
+CREATE UNIQUE INDEX "OutputEvaluation_modelOutputId_evaluationId_key" ON "OutputEvaluation"("modelOutputId", "evaluationId");
+
+-- AddForeignKey
+ALTER TABLE "OutputEvaluation" ADD CONSTRAINT "OutputEvaluation_modelOutputId_fkey" FOREIGN KEY ("modelOutputId") REFERENCES "ModelOutput"("id") ON DELETE CASCADE ON UPDATE CASCADE;
+
+-- AddForeignKey
+ALTER TABLE "OutputEvaluation" ADD CONSTRAINT "OutputEvaluation_evaluationId_fkey" FOREIGN KEY ("evaluationId") REFERENCES "Evaluation"("id") ON DELETE CASCADE ON UPDATE CASCADE;

prisma/schema.prisma

@@ -132,6 +132,7 @@ model ModelOutput {
 enum EvalType {
     CONTAINS
     DOES_NOT_CONTAIN
+    GPT4_EVAL
 }
 
 model Evaluation {

src/components/AutoResizeTextArea.tsx

@@ -4,7 +4,7 @@ import React from "react";
 
 export const AutoResizeTextarea: React.ForwardRefRenderFunction<
   HTMLTextAreaElement,
-  TextareaProps
+  TextareaProps & { minRows?: number }
 > = (props, ref) => {
   return (
     <Textarea

src/components/OutputsTable/EditEvaluations.tsx

@@ -11,12 +11,14 @@ import {
   FormLabel,
   Select,
   FormHelperText,
+  Code,
 } from "@chakra-ui/react";
 import { type Evaluation, EvalType } from "@prisma/client";
 import { useCallback, useState } from "react";
 import { BsPencil, BsX } from "react-icons/bs";
 import { api } from "~/utils/api";
 import { useExperiment, useHandledAsyncCallback } from "~/utils/hooks";
+import AutoResizeTextArea from "../AutoResizeTextArea";
 
 type EvalValues = Pick<Evaluation, "label" | "value" | "evalType">;
 
@@ -36,7 +38,7 @@ export function EvaluationEditor(props: {
     <VStack borderTopWidth={1} borderColor="gray.200" py={4}>
       <HStack w="100%">
         <FormControl flex={1}>
-          <FormLabel fontSize="sm">Evaluation Name</FormLabel>
+          <FormLabel fontSize="sm">Eval Name</FormLabel>
           <Input
             size="sm"
             value={values.label}
@@ -44,7 +46,7 @@ export function EvaluationEditor(props: {
           />
         </FormControl>
         <FormControl flex={1}>
-          <FormLabel fontSize="sm">Match Type</FormLabel>
+          <FormLabel fontSize="sm">Eval Type</FormLabel>
           <Select
             size="sm"
             value={values.evalType}
@@ -63,6 +65,7 @@ export function EvaluationEditor(props: {
           </Select>
         </FormControl>
       </HStack>
+      {["CONTAINS", "DOES_NOT_CONTAIN"].includes(values.evalType) && (
         <FormControl>
           <FormLabel fontSize="sm">Match String</FormLabel>
           <Input
@@ -71,9 +74,28 @@ export function EvaluationEditor(props: {
             onChange={(e) => setValues((values) => ({ ...values, value: e.target.value }))}
           />
           <FormHelperText>
-          This string will be interpreted as a regex and checked against each model output.
+            This string will be interpreted as a regex and checked against each model output. You
+            can include scenario variables using <Code>{"{{curly_braces}}"}</Code>
           </FormHelperText>
         </FormControl>
+      )}
+      {values.evalType === "GPT4_EVAL" && (
+        <FormControl pt={2}>
+          <FormLabel fontSize="sm">GPT4 Instructions</FormLabel>
+          <AutoResizeTextArea
+            size="sm"
+            value={values.value}
+            onChange={(e) => setValues((values) => ({ ...values, value: e.target.value }))}
+            minRows={3}
+          />
+          <FormHelperText>
+            Give instructions to GPT-4 for how to evaluate your prompt. It will have access to the
+            full scenario as well as the output it is evaluating. It will <strong>not</strong> have
+            access to the specific prompt variant, so be sure to be clear about the task you want it
+            to perform.
+          </FormHelperText>
+        </FormControl>
+      )}
       <HStack alignSelf="flex-end">
         <Button size="sm" onClick={props.onCancel} colorScheme="gray">
           Cancel

src/components/OutputsTable/EditScenarioVars.tsx

@@ -1,4 +1,4 @@
-import { Text, Button, HStack, Heading, Icon, Input, Stack, Code } from "@chakra-ui/react";
+import { Text, Button, HStack, Heading, Icon, Input, Stack } from "@chakra-ui/react";
 import { useState } from "react";
 import { BsCheck, BsX } from "react-icons/bs";
 import { api } from "~/utils/api";
@@ -36,8 +36,7 @@ export default function EditScenarioVars() {
       <Heading size="sm">Scenario Variables</Heading>
       <Stack spacing={2}>
         <Text fontSize="sm">
-          Scenario variables can be used in your prompt variants as well as evaluations. Reference
+          Scenario variables can be used in your prompt variants as well as evaluations.
-          them using <Code>{"{{curly_braces}}"}</Code>.
         </Text>
         <HStack spacing={0}>
           <Input

src/components/OutputsTable/OutputCell/OutputStats.tsx

@@ -2,7 +2,7 @@ import { type SupportedModel } from "~/server/types";
 import { type Scenario } from "../types";
 import { type RouterOutputs } from "~/utils/api";
 import { calculateTokenCost } from "~/utils/calculateTokenCost";
-import { HStack, Icon, Text } from "@chakra-ui/react";
+import { HStack, Icon, Text, Tooltip } from "@chakra-ui/react";
 import { BsCheck, BsClock, BsCurrencyDollar, BsX } from "react-icons/bs";
 import { CostTooltip } from "~/components/tooltip/CostTooltip";
 
@@ -36,7 +36,12 @@ export const OutputStats = ({
         {modelOutput.outputEvaluation.map((evaluation) => {
           const passed = evaluation.result > 0.5;
           return (
-            <HStack spacing={0} key={evaluation.id}>
+            <Tooltip
+              isDisabled={!evaluation.details}
+              label={evaluation.details}
+              key={evaluation.id}
+            >
+              <HStack spacing={0}>
                 <Text>{evaluation.evaluation.label}</Text>
                 <Icon
                   as={passed ? BsCheck : BsX}
@@ -44,6 +49,7 @@ export const OutputStats = ({
                   boxSize={6}
                 />
               </HStack>
+            </Tooltip>
           );
         })}
       </HStack>

src/server/utils/evaluations.ts

@@ -4,7 +4,7 @@ import { runOneEval } from "./runOneEval";
 import { type Scenario } from "~/components/OutputsTable/types";
 
 const saveResult = async (evaluation: Evaluation, scenario: Scenario, modelOutput: ModelOutput) => {
-  const result = runOneEval(evaluation, scenario, modelOutput);
+  const result = await runOneEval(evaluation, scenario, modelOutput);
   return await prisma.outputEvaluation.upsert({
     where: {
       modelOutputId_evaluationId: {
@@ -15,10 +15,10 @@ const saveResult = async (evaluation: Evaluation, scenario: Scenario, modelOutpu
     create: {
       modelOutputId: modelOutput.id,
       evaluationId: evaluation.id,
-      result,
+      ...result,
     },
     update: {
-      result,
+      ...result,
     },
   });
 };
@@ -35,43 +35,6 @@ export const runEvalsForOutput = async (
   await Promise.all(
     evaluations.map(async (evaluation) => await saveResult(evaluation, scenario, modelOutput)),
   );
-
-  // const cells = await prisma.scenarioVariantCell.findMany({
-  //   where: {
-  //     promptVariantId: variantId,
-  //     retrievalStatus: "COMPLETE",
-  //     testScenario: { visible: true },
-  //   },
-  //   include: { testScenario: true, modelOutput: { include: { OutputEvaluation: true } } },
-  // });
-
-  // await Promise.all(
-  //   evaluations.map(async (evaluation) => {
-  //     const passCount = cells.filter((cell) =>
-  //       runOneEval(cell.modelOutput as ModelOutput, cell.testScenario, evaluation),
-  //     ).length;
-  //     const failCount = cells.length - passCount;
-
-  //     await prisma.evaluationResult.upsert({
-  //       where: {
-  //         evaluationId_promptVariantId: {
-  //           evaluationId: evaluation.id,
-  //           promptVariantId: variantId,
-  //         },
-  //       },
-  //       create: {
-  //         evaluationId: evaluation.id,
-  //         promptVariantId: variantId,
-  //         passCount,
-  //         failCount,
-  //       },
-  //       update: {
-  //         passCount,
-  //         failCount,
-  //       },
-  //     });
-  //   }),
-  // );
 };
 
 export const runAllEvals = async (experimentId: string) => {
@@ -113,42 +76,4 @@ export const runAllEvals = async (experimentId: string) => {
       );
     }),
   );
-
-  // const cells = await prisma.scenarioVariantCell.findMany({
-  //   where: {
-  //     promptVariantId: { in: variants.map((v) => v.id) },
-  //     testScenario: { visible: true },
-  //     statusCode: { notIn: [429] },
-  //   },
-  //   include: { testScenario: true, modelOutput: true },
-  // });
-
-  // await Promise.all(
-  //   variants.map(async (variant) => {
-  //     const variantCells = cells.filter((cell) => cell.promptVariantId === variant.id);
-  //     const passCount = variantCells.filter((cell) =>
-  //       runOneEval(cell.modelOutput as ModelOutput, cell.testScenario, evaluation),
-  //     ).length;
-  //     const failCount = variantCells.length - passCount;
-
-  //     await prisma.evaluationResult.upsert({
-  //       where: {
-  //         evaluationId_promptVariantId: {
-  //           evaluationId: evaluation.id,
-  //           promptVariantId: variant.id,
-  //         },
-  //       },
-  //       create: {
-  //         evaluationId: evaluation.id,
-  //         promptVariantId: variant.id,
-  //         passCount,
-  //         failCount,
-  //       },
-  //       update: {
-  //         passCount,
-  //         failCount,
-  //       },
-  //     });
-  //   }),
-  // );
 };

src/server/utils/runOneEval.ts

@@ -1,32 +1,93 @@
 import { type Evaluation, type ModelOutput, type TestScenario } from "@prisma/client";
 import { type ChatCompletion } from "openai/resources/chat";
 import { type VariableMap, fillTemplate } from "./fillTemplate";
+import { openai } from "./openai";
+import dedent from "dedent";
 
-export const runOneEval = (
+export const runGpt4Eval = async (
+  evaluation: Evaluation,
+  scenario: TestScenario,
+  message: ChatCompletion.Choice.Message,
+): Promise<{ result: number; details: string }> => {
+  const output = await openai.chat.completions.create({
+    model: "gpt-4-0613",
+    messages: [
+      {
+        role: "system",
+        content: dedent`
+        You are a highly intelligent AI model and have been tasked with evaluating the quality of a simpler model. Your objective is to determine whether the simpler model has produced a successful and correct output. You should return "true" if the output was successful and "false" if it was not. Pay more attention to the semantics of the output than the formatting. Success is defined in the following terms:
+        ---
+        ${evaluation.value}
+        `,
+      },
+      {
+        role: "user",
+        content: `Scenario:\n---\n${JSON.stringify(scenario.variableValues, null, 2)}`,
+      },
+      {
+        role: "user",
+        content: `The full output of the simpler message:\n---\n${JSON.stringify(
+          message.content ?? message.function_call,
+          null,
+          2,
+        )}`,
+      },
+    ],
+    function_call: {
+      name: "report_success",
+    },
+    functions: [
+      {
+        name: "report_success",
+        parameters: {
+          type: "object",
+          required: ["thoughts", "success"],
+          properties: {
+            thoughts: {
+              type: "string",
+              description: "Explain your reasoning for considering this a pass or fail",
+            },
+            success: {
+              type: "boolean",
+              description:
+                "Whether the simpler model successfully completed the task for this scenario",
+            },
+          },
+        },
+      },
+    ],
+  });
+
+  try {
+    const out = JSON.parse(output.choices[0]?.message?.function_call?.arguments ?? "");
+    return { result: out.success ? 1 : 0, details: out.thoughts ?? JSON.stringify(out) };
+  } catch (e) {
+    console.error(e);
+    return { result: 0, details: "Error parsing GPT-4 output" };
+  }
+};
+
+export const runOneEval = async (
   evaluation: Evaluation,
   scenario: TestScenario,
   modelOutput: ModelOutput,
-): number => {
+): Promise<{ result: number; details?: string }> => {
   const output = modelOutput.output as unknown as ChatCompletion;
 
   const message = output?.choices?.[0]?.message;
 
-  if (!message) return 0;
+  if (!message) return { result: 0 };
 
   const stringifiedMessage = message.content ?? JSON.stringify(message.function_call);
 
   const matchRegex = fillTemplate(evaluation.value, scenario.variableValues as VariableMap);
 
-  let result;
-
   switch (evaluation.evalType) {
     case "CONTAINS":
-      result = stringifiedMessage.match(matchRegex) !== null ? 1 : 0;
+      return { result: stringifiedMessage.match(matchRegex) !== null ? 1 : 0 };
-      break;
     case "DOES_NOT_CONTAIN":
-      result = stringifiedMessage.match(matchRegex) === null ? 1 : 0;
+      return { result: stringifiedMessage.match(matchRegex) === null ? 1 : 0 };
-      break;
+    case "GPT4_EVAL":
+      return await runGpt4Eval(evaluation, scenario, message);
   }
-
-  return result;
 };