new parser for json + thinking output + tests using deepseek r1 on gr…

…oq (#195)

.changeset/little-cougars-bow.md

@@ -0,0 +1,5 @@
+---
+"@instructor-ai/instructor": minor
+---
+
+adding a new mode to support parsing thinking blocks out of markdown json responses (R1)

.github/workflows/test-pr.yml

@@ -23,6 +23,7 @@ jobs:
       TOGETHER_API_KEY: ${{ secrets.TOGETHER_API_KEY }}
       ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
       GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
+      GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
     steps:
       - uses: actions/checkout@v3
         with:

.github/workflows/test.yml

@@ -15,7 +15,7 @@ jobs:
       ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
       TOGETHER_API_KEY: ${{ secrets.TOGETHER_API_KEY }}
       GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
-
+      GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
     steps:
       - uses: actions/checkout@v3
         with:

package.json

@@ -44,14 +44,17 @@
     "outputs",
     "zod"
   ],
-  "author": "Jason Liu",
+  "contributors": [
+    "Dimitri Kennedy",
+    "Jason Liu"
+  ],
   "license": "MIT",
   "bugs": {
     "url": "https://github.com/instructor-ai/instructor-js/issues"
   },
   "homepage": "https://github.com/instructor-ai/instructor-js#readme",
   "dependencies": {
-    "zod-stream": "2.0.0",
+    "zod-stream": "3.0.0",
     "zod-validation-error": "^3.4.0"
   },
   "peerDependencies": {

src/constants/providers.ts

@@ -1,9 +1,16 @@
 import { omit } from "@/lib"
 import OpenAI from "openai"
 import { z } from "zod"
-import { withResponseModel, MODE as ZMODE, type Mode } from "zod-stream"
+import { thinkingJsonParser, withResponseModel, MODE as ZMODE } from "zod-stream"
+
+import { Mode } from "../types"
+
+export const MODE: typeof ZMODE = ZMODE
+
+export const MODE_TO_RESPONSE_PARSER = {
+  [MODE.THINKING_MD_JSON]: thinkingJsonParser
+}
 
-export const MODE = ZMODE
 export const PROVIDERS = {
   OAI: "OAI",
   ANYSCALE: "ANYSCALE",
@@ -12,6 +19,7 @@ export const PROVIDERS = {
   GROQ: "GROQ",
   OTHER: "OTHER"
 } as const
+
 export type Provider = keyof typeof PROVIDERS
 
 export const PROVIDER_SUPPORTED_MODES: {
@@ -98,7 +106,8 @@ export const PROVIDER_SUPPORTED_MODES_BY_MODEL = {
     [MODE.TOOLS]: ["*"],
     [MODE.JSON]: ["*"],
     [MODE.MD_JSON]: ["*"],
-    [MODE.JSON_SCHEMA]: ["*"]
+    [MODE.JSON_SCHEMA]: ["*"],
+    [MODE.THINKING_MD_JSON]: ["*"]
   },
   [PROVIDERS.OAI]: {
     [MODE.FUNCTIONS]: ["*"],
@@ -122,6 +131,7 @@ export const PROVIDER_SUPPORTED_MODES_BY_MODEL = {
   },
   [PROVIDERS.GROQ]: {
     [MODE.TOOLS]: ["*"],
-    [MODE.MD_JSON]: ["*"]
+    [MODE.MD_JSON]: ["*"],
+    [MODE.THINKING_MD_JSON]: ["deepseek-r1-distill-llama-70b"]
   }
 }

src/instructor.ts

@@ -11,18 +11,19 @@ import {
 import OpenAI from "openai"
 import { Stream } from "openai/streaming"
 import { z, ZodError } from "zod"
-import ZodStream, { OAIResponseParser, OAIStream, withResponseModel, type Mode } from "zod-stream"
+import ZodStream, { OAIResponseParser, OAIStream, withResponseModel } from "zod-stream"
 import { fromZodError } from "zod-validation-error"
 
 import {
+  MODE_TO_RESPONSE_PARSER,
   NON_OAI_PROVIDER_URLS,
   Provider,
   PROVIDER_PARAMS_TRANSFORMERS,
   PROVIDER_SUPPORTED_MODES,
   PROVIDERS
 } from "./constants/providers"
 import { iterableTee } from "./lib"
-import { ClientTypeChatCompletionParams, CompletionMeta } from "./types"
+import { ClientTypeChatCompletionParams, CompletionMeta, Mode } from "./types"
 
 const MAX_RETRIES_DEFAULT = 0
 
@@ -68,6 +69,7 @@ class Instructor<C> {
         : this.client?.baseURL.includes(NON_OAI_PROVIDER_URLS.TOGETHER) ? PROVIDERS.TOGETHER
         : this.client?.baseURL.includes(NON_OAI_PROVIDER_URLS.OAI) ? PROVIDERS.OAI
         : this.client?.baseURL.includes(NON_OAI_PROVIDER_URLS.ANTHROPIC) ? PROVIDERS.ANTHROPIC
+        : this.client?.baseURL.includes(NON_OAI_PROVIDER_URLS.GROQ) ? PROVIDERS.GROQ
         : PROVIDERS.OTHER
       : PROVIDERS.OTHER
 
@@ -187,13 +189,22 @@ class Instructor<C> {
         throw error
       }
 
-      const parsedCompletion = OAIResponseParser(
-        completion as OpenAI.Chat.Completions.ChatCompletion
-      )
+      const responseParser = MODE_TO_RESPONSE_PARSER?.[this.mode] ?? OAIResponseParser
+      const parsedCompletion = responseParser(completion as OpenAI.Chat.Completions.ChatCompletion)
 
       try {
-        const data = JSON.parse(parsedCompletion) as z.infer<T> & { _meta?: CompletionMeta }
-        return { ...data, _meta: { usage: completion?.usage ?? undefined } }
+        const responseJson = parsedCompletion.json ?? parsedCompletion
+        const data = JSON.parse(responseJson) as z.infer<T> & {
+          _meta?: CompletionMeta
+          thinking?: string
+        }
+        return {
+          ...data,
+          _meta: {
+            usage: completion?.usage ?? undefined,
+            thinking: parsedCompletion?.thinking ?? undefined
+          }
+        }
       } catch (error) {
         this.log(
           "error",

src/types/index.ts

@@ -61,9 +61,10 @@ export type LogLevel = "debug" | "info" | "warn" | "error"
 
 export type CompletionMeta = Partial<ZCompletionMeta> & {
   usage?: OpenAI.CompletionUsage
+  thinking?: string
 }
 
-export type Mode = ZMode
+export type Mode = ZMode | "THINKING_MD_JSON"
 
 export type ResponseModel<T extends z.AnyZodObject> = ZResponseModel<T>
 

tests/deepseek.test.ts

@@ -0,0 +1,80 @@
+import Instructor from "@/index"
+import { describe, expect, test } from "bun:test"
+import OpenAI from "openai"
+import { z } from "zod"
+
+const textBlock = `
+In our recent online meeting, participants from various backgrounds joined to discuss the upcoming tech conference. The names and contact details of the participants were as follows:
+
+- Name: John Doe, Email: [email protected], Twitter: @TechGuru44
+- Name: Jane Smith, Email: [email protected], Twitter: @DigitalDiva88
+- Name: Alex Johnson, Email: [email protected], Twitter: @CodeMaster2023
+
+During the meeting, we agreed on several key points. The conference will be held on March 15th, 2024, at the Grand Tech Arena located at 4521 Innovation Drive. Dr. Emily Johnson, a renowned AI researcher, will be our keynote speaker.
+
+The budget for the event is set at $50,000, covering venue costs, speaker fees, and promotional activities. Each participant is expected to contribute an article to the conference blog by February 20th.
+
+A follow-up meeting is scheduled for January 25th at 3 PM GMT to finalize the agenda and confirm the list of speakers.
+`
+
+const ExtractionValuesSchema = z.object({
+  users: z
+    .array(
+      z.object({
+        name: z.string(),
+        email: z.string(),
+        twitter: z.string()
+      })
+    )
+    .min(3),
+  conference: z.object({
+    date: z.string(),
+    venue: z.string(),
+    budget: z.number(),
+    keynoteSpeaker: z.string()
+  }),
+  nextMeeting: z.object({
+    date: z.string(),
+    time: z.string(),
+    timezone: z.string()
+  })
+})
+
+describe("thinking parser - live tests", () => {
+  test("should parse r1 response with thinking tags", async () => {
+    const groq = new OpenAI({
+      apiKey: process.env["GROQ_API_KEY"] ?? undefined,
+      baseURL: "https://api.groq.com/openai/v1"
+    })
+
+    const client = Instructor({
+      client: groq,
+      mode: "THINKING_MD_JSON",
+      debug: true
+    })
+
+    const result = await client.chat.completions.create({
+      messages: [{ role: "user", content: textBlock }],
+      model: "deepseek-r1-distill-llama-70b",
+      response_model: { schema: ExtractionValuesSchema, name: "Extract" },
+      max_retries: 4
+    })
+
+    console.log("result", result)
+
+    expect(result._meta?.thinking).toBeDefined()
+    expect(typeof result._meta?.thinking).toBe("string")
+
+    expect(result.users).toHaveLength(3)
+    expect(result.users[0]).toHaveProperty("name")
+    expect(result.users[0]).toHaveProperty("email")
+    expect(result.users[0]).toHaveProperty("twitter")
+
+    expect(result.conference).toBeDefined()
+    expect(result.conference.budget).toBe(50000)
+    expect(result.conference.keynoteSpeaker).toBe("Dr. Emily Johnson")
+
+    expect(result.nextMeeting).toBeDefined()
+    expect(result.nextMeeting.timezone).toBe("GMT")
+  })
+})

tests/mode.test.ts

@@ -128,23 +128,11 @@ describe("Modes", async () => {
   const testCases = createTestCases()
 
   for await (const { model, mode, provider, defaultMessage } of testCases) {
-    if (provider !== PROVIDERS.GROQ) {
-      test(`${provider}: Should return extracted name and age for model ${model} and mode ${mode}`, async () => {
-        const user = await extractUser(model, mode, provider, defaultMessage)
-
-        expect(user.name).toEqual("Jason Liu")
-        expect(user.age).toEqual(30)
-      })
-    } else {
-      test.todo(
-        `${provider}: Should return extracted name and age for model ${model} and mode ${mode}`,
-        async () => {
-          const user = await extractUser(model, mode, provider, defaultMessage)
-
-          expect(user.name).toEqual("Jason Liu")
-          expect(user.age).toEqual(30)
-        }
-      )
-    }
+    test(`${provider}: Should return extracted name and age for model ${model} and mode ${mode}`, async () => {
+      const user = await extractUser(model, mode, provider, defaultMessage)
+
+      expect(user.name).toEqual("Jason Liu")
+      expect(user.age).toEqual(30)
+    })
   }
 })