server/llm: update llm packages

Author: haraldschilly

src/.claude/settings.local.json

@@ -4,8 +4,15 @@
       "Bash(pnpm tsc:*)",
       "Bash(pnpm build:*)",
       "Bash(git add:*)",
-      "Bash(git commit:*)"
+      "Bash(git commit:*)",
+      "Bash(node:*)",
+      "Bash(grep:*)",
+      "Bash(find:*)",
+      "WebFetch(domain:github.com)",
+      "WebFetch(domain:cocalc.com)",
+      "WebFetch(domain:doc.cocalc.com)",
+      "Bash(npm show:*)"
     ],
     "deny": []
   }
-}
+}

src/packages/pnpm-lock.yaml

@@ -19,18 +19,11 @@ import { numTokens } from "./chatgpt-numtokens";
 const log = getLogger("llm:anthropic");
 
 function getModelName(model: AnthropicModel): string {
-  // The -4k and -8k variants have a limited context window (by us here) while offered for free
-  if (model === "claude-3-sonnet-4k") {
-    model = "claude-3-sonnet";
-  } else if (model === "claude-3-haiku-8k") {
-    model = "claude-3-haiku";
-  } else if (model === "claude-3-opus-8k") {
-    model = "claude-3-opus";
-  } else if (model === "claude-3-5-sonnet-4k") {
-    model = "claude-3-5-sonnet";
+  const id = ANTHROPIC_VERSION[model];
+  if (id == null) {
+    throw new Error(`Anthropic model ${model} is no longer supported`);
   }
-  // now we have a valid name, and we have to append their static version number
-  return `${model}-${ANTHROPIC_VERSION[model]}`;
+  return id;
 }
 
 interface AnthropicOpts {

src/packages/server/llm/anthropic.ts

@@ -110,6 +110,10 @@ export class GoogleGenAIClient {
       model: modelName,
       apiKey: this.apiKey,
       maxOutputTokens: maxTokens,
+      // Only enable thinking tokens for Gemini 2.5 models
+      ...(modelName === "gemini-2.5-flash" || modelName === "gemini-2.5-pro"
+        ? { maxReasoningTokens: 1024 }
+        : {}),
       streaming: true,
     });
 
@@ -155,8 +159,9 @@ export class GoogleGenAIClient {
       ...history.map(({ content }) => content),
     ]);
 
-    const { totalTokens: completion_tokens } =
-      await geminiPro.countTokens(output);
+    const { totalTokens: completion_tokens } = await geminiPro.countTokens(
+      output,
+    );
 
     log.debug("chatGemini successful", { prompt_tokens, completion_tokens });
 

src/packages/server/llm/google-genai-client.ts

@@ -112,6 +112,12 @@ test_llm("google")("Google GenAI", () => {
   test("gemini 2.0 flash lite works", async () => {
     llmGoogle("gemini-2.0-flash-lite-8k");
   });
+  test("gemini 2.5 flash works", async () => {
+    llmGoogle("gemini-2.5-flash-8k");
+  });
+  test("gemini 2.5 pro works", async () => {
+    llmGoogle("gemini-2.5-pro-8k");
+  });
 });
 
 test_llm("mistralai")("Mistral AI", () => {
@@ -129,12 +135,14 @@ test_llm("mistralai")("Mistral AI", () => {
 });
 
 test_llm("anthropic")("Anthropic", () => {
-  const haiku: AnthropicModel = "claude-3-haiku";
-  const sonnet: AnthropicModel = "claude-3-5-sonnet-4k";
-  const opus: AnthropicModel = "claude-3-opus-8k";
+  const haiku: AnthropicModel = "claude-3-5-haiku-8k";
+  const sonnet: AnthropicModel = "claude-4-sonnet-8k";
+  const opus: AnthropicModel = "claude-4-opus-8k";
 
   test("model", () => {
     expect(isAnthropicModel(haiku)).toBe(true);
+    expect(isAnthropicModel(sonnet)).toBe(true);
+    expect(isAnthropicModel(opus)).toBe(true);
   });
 
   test("haiku", async () => {

src/packages/server/llm/test/00.test.ts renamed to src/packages/server/llm/test/models.test.ts

@@ -55,12 +55,12 @@
     "@google-cloud/storage-transfer": "^3.3.0",
     "@google/generative-ai": "^0.14.0",
     "@isaacs/ttlcache": "^1.4.1",
-    "@langchain/anthropic": "^0.3.18",
-    "@langchain/core": "^0.3.46",
-    "@langchain/google-genai": "^0.2.4",
-    "@langchain/mistralai": "^0.2.0",
-    "@langchain/ollama": "^0.2.0",
-    "@langchain/openai": "^0.5.5",
+    "@langchain/anthropic": "^0.3.24",
+    "@langchain/core": "^0.3.64",
+    "@langchain/google-genai": "^0.2.15",
+    "@langchain/mistralai": "^0.2.1",
+    "@langchain/ollama": "^0.2.3",
+    "@langchain/openai": "^0.6.1",
     "@node-saml/passport-saml": "^5.0.1",
     "@passport-js/passport-twitter": "^1.0.8",
     "@passport-next/passport-google-oauth2": "^1.0.0",
@@ -97,7 +97,7 @@
     "nanoid": "^3.3.8",
     "node-zendesk": "^5.0.13",
     "nodemailer": "^6.9.16",
-    "openai": "^4.95.1",
+    "openai": "^4.104.0",
     "parse-domain": "^5.0.0",
     "passport": "^0.6.0",
     "passport-activedirectory": "^1.0.4",