frontend/llm: collect LLM descriptions in the universal llm-util

sagemathinc · Mar 15, 2024 · cb66d65 · cb66d65
1 parent 5d3f293
commit cb66d65
Show file tree

Hide file tree

Showing 3 changed files with 45 additions and 54 deletions.
diff --git a/src/packages/frontend/frame-editors/llm/model-switch.tsx b/src/packages/frontend/frame-editors/llm/model-switch.tsx
@@ -6,6 +6,7 @@ import { CSS, redux, useTypedRedux } from "@cocalc/frontend/app-framework";
 import { LanguageModelVendorAvatar } from "@cocalc/frontend/components/language-model-icon";
 import {
   DEFAULT_MODEL,
+  LLM_DESCR,
   LLM_USERNAMES,
   LanguageModel,
   MISTRAL_MODELS,
@@ -90,7 +91,7 @@ export default function ModelSwitch({
     );
     const tooltip = (
       <>
-        <strong>{model}</strong>: {title}
+        <strong>{model}</strong> – {title}
       </>
     );
     const display = (
@@ -108,64 +109,26 @@ export default function ModelSwitch({
   function appendOpenAI(ret: NonNullable<SelectProps["options"]>) {
     if (!showOpenAI) return null;
 
-    makeLLMOption(
-      ret,
-      "gpt-3.5-turbo",
-      "OpenAI's fastest model, great for most everyday tasks (4k token context)",
-    );
-    makeLLMOption(
-      ret,
-      "gpt-3.5-turbo-16k",
-      `Same as ${modelToName(
-        "gpt-3.5-turbo",
-      )} but with much larger context size (16k token context)`,
-    );
-    makeLLMOption(
-      ret,
-      "gpt-4",
-      "GPT-4 can follow complex instructions in natural language and solve difficult problems with accuracy. (8k token context)",
-    );
-    makeLLMOption(
-      ret,
-      "gpt-4-turbo-preview",
-      "GPT-4 Turbo is more powerful, has fresher knowledge and offered at a lower price  than GPT-4. (128k token context)",
-    );
+    makeLLMOption(ret, "gpt-3.5-turbo", LLM_DESCR["gpt-3.5-turbo"]);
+    makeLLMOption(ret, "gpt-3.5-turbo-16k", LLM_DESCR["gpt-3.5-turbo-16k"]);
+    makeLLMOption(ret, "gpt-4", LLM_DESCR["gpt-4"]);
+    makeLLMOption(ret, "gpt-4-turbo-preview", LLM_DESCR["gpt-4-turbo-preview"]);
   }
 
   function appendGoogle(ret: NonNullable<SelectProps["options"]>) {
     if (!showGoogle) return null;
 
-    return (
-      <>
-        {makeLLMOption(
-          ret,
-          GOOGLE_GEMINI,
-          `Google's Gemini Pro Generative AI model (30k token context)`,
-        )}
-      </>
-    );
+    return <>{makeLLMOption(ret, GOOGLE_GEMINI, LLM_DESCR[GOOGLE_GEMINI])}</>;
   }
 
   function appendMistral(ret: NonNullable<SelectProps["options"]>) {
     if (!showMistral) return null;
 
     return (
       <>
-        {makeLLMOption(
-          ret,
-          MISTRAL_MODELS[0],
-          "Fast, simple queries, short answers, less capabilities",
-        )}
-        {makeLLMOption(
-          ret,
-          MISTRAL_MODELS[1],
-          "Intermediate tasks, summarizing, generating documents, etc.",
-        )}
-        {makeLLMOption(
-          ret,
-          MISTRAL_MODELS[2],
-          "Slowest, most powerful, large reasoning capabilities",
-        )}
+        {makeLLMOption(ret, MISTRAL_MODELS[0], LLM_DESCR[MISTRAL_MODELS[0]])}
+        {makeLLMOption(ret, MISTRAL_MODELS[1], LLM_DESCR[MISTRAL_MODELS[1]])}
+        {makeLLMOption(ret, MISTRAL_MODELS[2], LLM_DESCR[MISTRAL_MODELS[2]])}
       </>
     );
   }
@@ -178,7 +141,8 @@ export default function ModelSwitch({
       const ollamaModel = toOllamaModel(key);
       const text = (
         <>
-          <strong>{display}</strong> {getPrice(ollamaModel)}: {desc ?? "Ollama"}
+          <strong>{display}</strong> {getPrice(ollamaModel)} –{" "}
+          {desc ?? "Ollama"}
         </>
       );
       ret.push({

diff --git a/src/packages/util/db-schema/llm-utils.ts b/src/packages/util/db-schema/llm-utils.ts
@@ -263,10 +263,7 @@ export function fromMistralService(model: MistralService) {
   return model.slice(MISTRAL_PREFIX.length);
 }
 
-// Map from psuedo account_id to what should be displayed to user.
-// This is used in various places in the frontend.
-// Google PaLM: https://cloud.google.com/vertex-ai/docs/generative-ai/pricing
-export const LLM_USERNAMES: {
+type LLM2String = {
   [key in
     | (typeof USER_SELECTABLE_LANGUAGE_MODELS)[number]
     | "chatgpt" // some additional ones, backwards compatibility
@@ -275,7 +272,12 @@ export const LLM_USERNAMES: {
     | "gpt-4-32k"
     | "text-bison-001"
     | "chat-bison-001"]: string;
-} = {
+};
+
+// Map from psuedo account_id to what should be displayed to user.
+// This is used in various places in the frontend.
+// Google PaLM: https://cloud.google.com/vertex-ai/docs/generative-ai/pricing
+export const LLM_USERNAMES: LLM2String = {
   chatgpt: "GPT-3.5",
   chatgpt3: "GPT-3.5",
   chatgpt4: "GPT-4",
@@ -292,6 +294,31 @@ export const LLM_USERNAMES: {
   "mistral-large-latest": "Mistral AI Large",
 } as const;
 
+// similar to the above, we map to short user-visible description texts
+// this comes next to the name, hence you do not have to mention the name
+export const LLM_DESCR: LLM2String = {
+  chatgpt: "Fast, great for everyday tasks. (OpenAI, 4k token context)",
+  chatgpt3: "Fast, great for everyday tasks. (OpenAI, 4k token context)",
+  chatgpt4:
+    "Can follow complex instructions and solve difficult problems. (OpenAI, 8k token context)",
+  "gpt-4":
+    "Can follow complex instructions and solve difficult problems. (OpenAI, 8k token context)",
+  "gpt-4-32k": "",
+  "gpt-3.5-turbo": "Fast, great for everyday tasks. (OpenAI, 4k token context)",
+  "gpt-3.5-turbo-16k": `Same as ${LLM_USERNAMES["gpt-3.5-turbo"]} but with larger 16k token context`,
+  "gpt-4-turbo-preview":
+    "More powerful, fresher knowledge, and lower price than GPT-4. (OpenAI, 128k token context)",
+  "text-bison-001": "",
+  "chat-bison-001": "",
+  "gemini-pro": "Google's Gemini Pro Generative AI model (30k token context)",
+  "mistral-small-latest":
+    "Fast, simple queries, short answers, less capabilities. (Mistral AI, 4k token context)",
+  "mistral-medium-latest":
+    "Intermediate tasks, summarizing, generating documents, etc. (Mistral AI, 4k token context)",
+  "mistral-large-latest":
+    "Most powerful, large reasoning capabilities, but slower. (Mistral AI, 4k token context)",
+} as const;
+
 export function isFreeModel(model: unknown) {
   if (isOllamaLLM(model)) return true;
   if (isMistralModel(model)) {

diff --git a/src/packages/util/db-schema/site-settings-extras.ts b/src/packages/util/db-schema/site-settings-extras.ts
@@ -282,7 +282,7 @@ export const EXTRAS: SettingsExtras = {
   },
   ollama_configuration: {
     name: "Ollama Configuration",
-    desc: 'Configure Ollama endpoints. e.g. Ollama has "gemma" installed and runs at localhost:11434: `{"gemma" : {"baseUrl": "http://localhost:11434/" , cocalc: {display: "Gemma", desc: "Google\'s Gemma Model"}}',
+    desc: 'Configure Ollama endpoints. e.g. Ollama has "gemma" installed and is available at localhost:11434: `{"gemma" : {"baseUrl": "http://localhost:11434/" , cocalc: {display: "Gemma", desc: "Google\'s Gemma Model"}}',
     default: "",
     multiline: 5,
     show: ollama_enabled,