feat: Allow llm model to be configurable

package.json

@@ -36,9 +36,9 @@
   "dependencies": {
     "@arcadeai/arcadejs": "^0.1.2",
     "@googleapis/youtube": "^20.0.0",
-    "@langchain/anthropic": "^0.3.9",
+    "@langchain/anthropic": "^0.3.11",
     "@langchain/community": "^0.3.22",
-    "@langchain/core": "^0.3.22",
+    "@langchain/core": "^0.3.27",
     "@langchain/google-vertexai-web": "^0.1.2",
     "@langchain/langgraph": "^0.2.31",
     "@langchain/langgraph-sdk": "^0.0.31",
@@ -54,6 +54,7 @@
     "express-session": "^1.18.1",
     "file-type": "^19.6.0",
     "google-auth-library": "^9.15.0",
+    "langchain": "^0.3.10",
     "langsmith": "0.2.15-rc.2",
     "moment": "^2.30.1",
     "passport": "^0.7.0",

scripts/generate-demo-post.ts

@@ -1,9 +1,5 @@
 import "dotenv/config";
 import { Client } from "@langchain/langgraph-sdk";
-// import {
-//   LINKEDIN_USER_ID,
-//   TWITTER_USER_ID,
-// } from "../src/agents/generate-post/constants.js";
 
 /**
  * Generate a post based on the Open Canvas project.
@@ -22,13 +18,6 @@ async function invokeGraph() {
     input: {
       links: [link],
     },
-    config: {
-      configurable: {
-        // By default, the graph will read these values from the environment
-        // [TWITTER_USER_ID]: process.env.TWITTER_USER_ID,
-        // [LINKEDIN_USER_ID]: process.env.LINKEDIN_USER_ID,
-      },
-    },
   });
 }
 

src/agents/generate-post/constants.ts

@@ -95,3 +95,5 @@ export const TWITTER_USER_ID = "twitterUserId";
 export const TWITTER_TOKEN = "twitterToken";
 export const TWITTER_TOKEN_SECRET = "twitterTokenSecret";
 export const INGEST_TWITTER_USERNAME = "ingestTwitterUsername";
+
+export const LLM_MODEL_NAME = "llmModel";

src/agents/generate-post/generate-post-state.ts

@@ -1,6 +1,6 @@
 import { Annotation, END } from "@langchain/langgraph";
 import { IngestDataAnnotation } from "../ingest-data/ingest-data-state.js";
-import { POST_TO_LINKEDIN_ORGANIZATION } from "./constants.js";
+import { LLM_MODEL_NAME, POST_TO_LINKEDIN_ORGANIZATION } from "./constants.js";
 import { DateType } from "../types.js";
 
 export type LangChainProduct = "langchain" | "langgraph" | "langsmith";
@@ -102,4 +102,14 @@ export const GeneratePostConfigurableAnnotation = Annotation.Root({
    * If true, [LINKEDIN_ORGANIZATION_ID] is required.
    */
   [POST_TO_LINKEDIN_ORGANIZATION]: Annotation<boolean | undefined>,
+  /**
+   * The name of the LLM to use for generations
+   * @default "gemini-2.0-flash-exp"
+   */
+  [LLM_MODEL_NAME]: Annotation<
+    "gemini-2.0-flash-exp" | "claude-3-5-sonnet-latest" | undefined
+  >({
+    reducer: (_state, update) => update,
+    default: () => "gemini-2.0-flash-exp",
+  }),
 });

src/agents/generate-post/nodes/condense-post.ts

@@ -1,8 +1,8 @@
-import { ChatAnthropic } from "@langchain/anthropic";
 import { GeneratePostAnnotation } from "../generate-post-state.js";
 import { STRUCTURE_INSTRUCTIONS, RULES } from "./geterate-post/prompts.js";
 import { parseGeneration } from "./geterate-post/utils.js";
-import { removeUrls } from "../../utils.js";
+import { getModelFromConfig, removeUrls } from "../../utils.js";
+import { LangGraphRunnableConfig } from "@langchain/langgraph";
 
 const CONDENSE_POST_PROMPT = `You're a highly skilled marketer at LangChain, working on crafting thoughtful and engaging content for LangChain's LinkedIn and Twitter pages.
 You wrote a post for the LangChain LinkedIn and Twitter pages, however it's a bit too long for Twitter, and thus needs to be condensed.
@@ -52,6 +52,7 @@ Follow all rules and instructions outlined above. The user message below will pr
  */
 export async function condensePost(
   state: typeof GeneratePostAnnotation.State,
+  config: LangGraphRunnableConfig,
 ): Promise<Partial<typeof GeneratePostAnnotation.State>> {
   if (!state.post) {
     throw new Error("No post found");
@@ -72,8 +73,7 @@ export async function condensePost(
     .replace("{link}", state.relevantLinks[0])
     .replace("{originalPostLength}", originalPostLength);
 
-  const condensePostModel = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
+  const condensePostModel = await getModelFromConfig(config, {
     temperature: 0.5,
   });
 

src/agents/generate-post/nodes/generate-content-report.ts

@@ -1,7 +1,7 @@
 import { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { GeneratePostAnnotation } from "../generate-post-state.js";
 import { LANGCHAIN_PRODUCTS_CONTEXT } from "../prompts.js";
-import { ChatAnthropic } from "@langchain/anthropic";
+import { getModelFromConfig } from "../../utils.js";
 
 const GENERATE_REPORT_PROMPT = `You are a highly regarded marketing employee at LangChain.
 You have been tasked with writing a marketing report on content submitted to you from a third party which uses LangChain's products.
@@ -85,10 +85,9 @@ ${pageContents.map((content, index) => `<Content index={${index + 1}}>\n${conten
 
 export async function generateContentReport(
   state: typeof GeneratePostAnnotation.State,
-  _config: LangGraphRunnableConfig,
+  config: LangGraphRunnableConfig,
 ): Promise<Partial<typeof GeneratePostAnnotation.State>> {
-  const reportModel = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
+  const reportModel = await getModelFromConfig(config, {
     temperature: 0,
   });
 

src/agents/generate-post/nodes/geterate-post/index.ts

@@ -1,11 +1,11 @@
 import { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { GeneratePostAnnotation } from "../../generate-post-state.js";
-import { ChatAnthropic } from "@langchain/anthropic";
 import { GENERATE_POST_PROMPT, REFLECTIONS_PROMPT } from "./prompts.js";
 import { formatPrompt, parseGeneration } from "./utils.js";
 import { ALLOWED_TIMES } from "../../constants.js";
 import { getReflections, RULESET_KEY } from "../../../../utils/reflections.js";
 import { getNextSaturdayDate } from "../../../../utils/date.js";
+import { getModelFromConfig } from "../../../utils.js";
 
 export async function generatePost(
   state: typeof GeneratePostAnnotation.State,
@@ -17,8 +17,7 @@ export async function generatePost(
   if (state.relevantLinks.length === 0) {
     throw new Error("No relevant links found");
   }
-  const postModel = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
+  const postModel = await getModelFromConfig(config, {
     temperature: 0.5,
   });
 

src/agents/generate-post/nodes/human-node/index.ts

@@ -78,7 +78,7 @@ Here is the report that was generated for the posts:\n${report}
 
 export async function humanNode(
   state: typeof GeneratePostAnnotation.State,
-  _config: LangGraphRunnableConfig,
+  config: LangGraphRunnableConfig,
 ): Promise<Partial<typeof GeneratePostAnnotation.State>> {
   if (!state.post) {
     throw new Error("No post found");
@@ -153,6 +153,7 @@ export async function humanNode(
       post: state.post,
       dateOrPriority: defaultDateString,
       userResponse: response.args,
+      config,
     });
 
     if (route === "rewrite_post") {

src/agents/generate-post/nodes/human-node/route-response.ts

@@ -1,5 +1,6 @@
-import { ChatAnthropic } from "@langchain/anthropic";
 import { z } from "zod";
+import { getModelFromConfig } from "../../../utils.js";
+import { LangGraphRunnableConfig } from "@langchain/langgraph";
 
 const ROUTE_RESPONSE_PROMPT = `You are an AI assistant tasked with routing a user's response to one of two possible routes based on their intention. The two possible routes are:
 
@@ -55,15 +56,16 @@ interface RouteResponseArgs {
   post: string;
   dateOrPriority: string;
   userResponse: string;
+  config: LangGraphRunnableConfig;
 }
 
 export async function routeResponse({
   post,
   dateOrPriority,
   userResponse,
+  config,
 }: RouteResponseArgs) {
-  const model = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
+  const model = await getModelFromConfig(config, {
     temperature: 0,
   });
 

src/agents/generate-post/nodes/rewrite-post.ts

@@ -1,7 +1,7 @@
 import { Client } from "@langchain/langgraph-sdk";
 import { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { GeneratePostAnnotation } from "../generate-post-state.js";
-import { ChatAnthropic } from "@langchain/anthropic";
+import { getModelFromConfig } from "../../utils.js";
 
 const REWRITE_POST_PROMPT = `You're a highly regarded marketing employee at LangChain, working on crafting thoughtful and engaging content for LangChain's LinkedIn and Twitter pages.
 You wrote a post for the LangChain LinkedIn and Twitter pages, however your boss has asked for some changes to be made before it can be published.
@@ -44,7 +44,7 @@ async function runReflections({
 
 export async function rewritePost(
   state: typeof GeneratePostAnnotation.State,
-  _config: LangGraphRunnableConfig,
+  config: LangGraphRunnableConfig,
 ): Promise<Partial<typeof GeneratePostAnnotation.State>> {
   if (!state.post) {
     throw new Error("No post found");
@@ -53,8 +53,7 @@ export async function rewritePost(
     throw new Error("No user response found");
   }
 
-  const rewritePostModel = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
+  const rewritePostModel = await getModelFromConfig(config, {
     temperature: 0.5,
   });
 

src/agents/generate-post/nodes/schedule-post/index.ts

@@ -1,7 +1,10 @@
 import { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { GeneratePostAnnotation } from "../../generate-post-state.js";
 import { Client } from "@langchain/langgraph-sdk";
-import { POST_TO_LINKEDIN_ORGANIZATION } from "../../constants.js";
+import {
+  LLM_MODEL_NAME,
+  POST_TO_LINKEDIN_ORGANIZATION,
+} from "../../constants.js";
 import { getScheduledDateSeconds } from "./find-date.js";
 import { SlackClient } from "../../../../clients/slack.js";
 import { getFutureDate } from "./get-future-date.js";
@@ -40,6 +43,8 @@ export async function schedulePost(
         [POST_TO_LINKEDIN_ORGANIZATION]:
           config.configurable?.[POST_TO_LINKEDIN_ORGANIZATION] ||
           process.env.POST_TO_LINKEDIN_ORGANIZATION,
+        [LLM_MODEL_NAME]:
+          config.configurable?.[LLM_MODEL_NAME] || "gemini-2.0-flash-exp",
       },
     },
     afterSeconds,

src/agents/generate-post/nodes/update-scheduled-date.ts

@@ -1,9 +1,10 @@
 import { z } from "zod";
 import { GeneratePostAnnotation } from "../generate-post-state.js";
-import { ChatAnthropic } from "@langchain/anthropic";
 import { toZonedTime } from "date-fns-tz";
 import { DateType } from "../../types.js";
 import { timezoneToUtc } from "../../../utils/date.js";
+import { getModelFromConfig } from "../../utils.js";
+import { LangGraphRunnableConfig } from "@langchain/langgraph";
 
 const SCHEDULE_POST_DATE_PROMPT = `You're an intelligent AI assistant tasked with extracting the date to schedule a social media post from the user's message.
 
@@ -40,14 +41,16 @@ const scheduleDateSchema = z.object({
 
 export async function updateScheduledDate(
   state: typeof GeneratePostAnnotation.State,
+  config: LangGraphRunnableConfig,
 ): Promise<Partial<typeof GeneratePostAnnotation.State>> {
   if (!state.userResponse) {
     throw new Error("No user response found");
   }
-  const model = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
-    temperature: 0.5,
-  }).withStructuredOutput(scheduleDateSchema, {
+  const model = (
+    await getModelFromConfig(config, {
+      temperature: 0.5,
+    })
+  ).withStructuredOutput(scheduleDateSchema, {
     name: "scheduleDate",
   });
   const pstDate = toZonedTime(new Date(), "America/Los_Angeles");

src/agents/ingest-data/ingest-data-graph.ts

@@ -10,7 +10,10 @@ import {
 } from "./ingest-data-state.js";
 import { ingestSlackData } from "./nodes/ingest-slack.js";
 import { Client } from "@langchain/langgraph-sdk";
-import { POST_TO_LINKEDIN_ORGANIZATION } from "../generate-post/constants.js";
+import {
+  LLM_MODEL_NAME,
+  POST_TO_LINKEDIN_ORGANIZATION,
+} from "../generate-post/constants.js";
 import { getUrlType } from "../utils.js";
 
 /**
@@ -69,6 +72,8 @@ async function generatePostFromMessages(
       config: {
         configurable: {
           [POST_TO_LINKEDIN_ORGANIZATION]: postToLinkedInOrg,
+          [LLM_MODEL_NAME]:
+            config.configurable?.[LLM_MODEL_NAME] || "gemini-2.0-flash-exp",
         },
       },
       afterSeconds,

src/agents/ingest-data/ingest-data-state.ts

@@ -1,6 +1,9 @@
 import { Annotation } from "@langchain/langgraph";
 import { SimpleSlackMessage } from "../../clients/slack.js";
-import { POST_TO_LINKEDIN_ORGANIZATION } from "../generate-post/constants.js";
+import {
+  LLM_MODEL_NAME,
+  POST_TO_LINKEDIN_ORGANIZATION,
+} from "../generate-post/constants.js";
 
 export type LangChainProduct = "langchain" | "langgraph" | "langsmith";
 export type SimpleSlackMessageWithLinks = SimpleSlackMessage & {
@@ -50,4 +53,14 @@ export const IngestDataConfigurableAnnotation = Annotation.Root({
    * If true, [LINKEDIN_ORGANIZATION_ID] is required.
    */
   [POST_TO_LINKEDIN_ORGANIZATION]: Annotation<boolean | undefined>,
+  /**
+   * The name of the LLM to use for generations
+   * @default "gemini-2.0-flash-exp"
+   */
+  [LLM_MODEL_NAME]: Annotation<
+    "gemini-2.0-flash-exp" | "claude-3-5-sonnet-latest" | undefined
+  >({
+    reducer: (_state, update) => update,
+    default: () => "gemini-2.0-flash-exp",
+  }),
 });

src/agents/reflection/index.ts

@@ -6,13 +6,13 @@ import {
   StateGraph,
 } from "@langchain/langgraph";
 import { z } from "zod";
-import { ChatAnthropic } from "@langchain/anthropic";
 import {
   getReflections,
   putReflections,
   RULESET_KEY,
 } from "../../utils/reflections.js";
 import { REFLECTION_PROMPT, UPDATE_RULES_PROMPT } from "./prompts.js";
+import { getModelFromConfig } from "../utils.js";
 
 const newRuleSchema = z.object({
   newRule: z.string().describe("The new rule to create."),
@@ -47,8 +47,7 @@ async function reflection(
   if (!config.store) {
     throw new Error("No store provided");
   }
-  const model = new ChatAnthropic({
-    model: "claude-3-5-sonnet-latest",
+  const model = await getModelFromConfig(config, {
     temperature: 0,
   });
 

src/agents/shared/nodes/verify-general.ts

@@ -1,12 +1,11 @@
 import { LangGraphRunnableConfig } from "@langchain/langgraph";
 import { GeneratePostAnnotation } from "../../generate-post/generate-post-state.js";
 import { z } from "zod";
-import { ChatAnthropic } from "@langchain/anthropic";
 import { FireCrawlLoader } from "@langchain/community/document_loaders/web/firecrawl";
 import { LANGCHAIN_PRODUCTS_CONTEXT } from "../../generate-post/prompts.js";
 import { VerifyContentAnnotation } from "../shared-state.js";
 import { RunnableLambda } from "@langchain/core/runnables";
-import { getPageText } from "../../utils.js";
+import { getModelFromConfig, getPageText } from "../../utils.js";
 
 type VerifyGeneralContentReturn = {
   relevantLinks: (typeof GeneratePostAnnotation.State)["relevantLinks"];
@@ -87,11 +86,13 @@ export async function getUrlContents(url: string): Promise<UrlContents> {
 
 export async function verifyGeneralContentIsRelevant(
   content: string,
+  config: LangGraphRunnableConfig,
 ): Promise<boolean> {
-  const relevancyModel = new ChatAnthropic({
-    model: "claude-3-5-sonnet-20241022",
-    temperature: 0,
-  }).withStructuredOutput(RELEVANCY_SCHEMA, {
+  const relevancyModel = (
+    await getModelFromConfig(config, {
+      temperature: 0,
+    })
+  ).withStructuredOutput(RELEVANCY_SCHEMA, {
     name: "relevancy",
   });
 
@@ -125,14 +126,17 @@ export async function verifyGeneralContentIsRelevant(
  */
 export async function verifyGeneralContent(
   state: typeof VerifyContentAnnotation.State,
-  _config: LangGraphRunnableConfig,
+  config: LangGraphRunnableConfig,
 ): Promise<VerifyGeneralContentReturn> {
   const urlContents = await new RunnableLambda<string, UrlContents>({
     func: getUrlContents,
   })
     .withConfig({ runName: "get-url-contents" })
     .invoke(state.link);
-  const relevant = await verifyGeneralContentIsRelevant(urlContents.content);
+  const relevant = await verifyGeneralContentIsRelevant(
+    urlContents.content,
+    config,
+  );
 
   if (relevant) {
     return {