browser/interfaces: add logprobs

README.md

@@ -134,6 +134,8 @@ ollama.chat(request)
   - `format` `<string>`: (Optional) Set the expected format of the response (`json`).
   - `stream` `<boolean>`: (Optional) When true an `AsyncGenerator` is returned.
   - `think` `<boolean | "high" | "medium" | "low">`: (Optional) Enable model thinking. Use `true`/`false` or specify a level. Requires model support.
+  - `logprobs` `<boolean>`: (Optional) Return log probabilities for tokens. Requires model support.
+  - `top_logprobs` `<number>`: (Optional) Number of top log probabilities to return per token when `logprobs` is enabled.
   - `keep_alive` `<string | number>`: (Optional) How long to keep the model loaded. A number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc.)
   - `tools` `<Tool[]>`: (Optional) A list of tool calls the model may make.
   - `options` `<Options>`: (Optional) Options to configure the runtime.
@@ -157,6 +159,8 @@ ollama.generate(request)
   - `format` `<string>`: (Optional) Set the expected format of the response (`json`).
   - `stream` `<boolean>`: (Optional) When true an `AsyncGenerator` is returned.
   - `think` `<boolean | "high" | "medium" | "low">`: (Optional) Enable model thinking. Use `true`/`false` or specify a level. Requires model support.
+  - `logprobs` `<boolean>`: (Optional) Return log probabilities for tokens. Requires model support.
+  - `top_logprobs` `<number>`: (Optional) Number of top log probabilities to return per token when `logprobs` is enabled.
   - `keep_alive` `<string | number>`: (Optional) How long to keep the model loaded. A number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc.)
   - `options` `<Options>`: (Optional) Options to configure the runtime.
 - Returns: `<GenerateResponse>`

examples/logprobs/logprobs-chat.ts

@@ -0,0 +1,31 @@
+import { Ollama, Logprob } from 'ollama';
+
+function printLogprobs(entries: Logprob[], label: string) {
+  console.log(`\n${label}:`)
+  for (const entry of entries) {
+    console.log(`  token=${entry.token.padEnd(12)} logprob=${entry.logprob.toFixed(3)}`)
+    for (const alt of entry.top_logprobs ?? []) {
+      console.log(`    alt -> ${alt.token.padEnd(12)} (${alt.logprob.toFixed(3)})`)
+    }
+  }
+}
+
+async function main() {
+  const client = new Ollama()
+  console.log(`Using model: gemma3`)
+
+  const chatResponse = await client.chat({
+    model: 'gemma3',
+    messages: [{ role: 'user', content: 'Say hello in one word.' }],
+    logprobs: true,
+    top_logprobs: 3,
+  })
+  console.log('Chat response:', chatResponse.message.content)
+  printLogprobs(chatResponse.logprobs ?? [], 'chat logprobs')
+}
+
+main().catch((err) => {
+  console.error(err)
+  process.exitCode = 1
+})
+

src/interfaces.ts

@@ -57,6 +57,8 @@
   images?: Uint8Array[] | string[]
   keep_alive?: string | number // a number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc)
   think?: boolean | 'high' | 'medium' | 'low'
+  logprobs?: boolean
+  top_logprobs?: number
 
   options?: Partial<Options>
 }
@@ -74,7 +76,7 @@
   function: {
     name: string;
     arguments: {
       [key: string]: any;
     };
   };
 }
@@ -87,15 +89,15 @@
     type?: string;
     parameters?: {
       type?: string;
       $defs?: any;
       items?: any;
       required?: string[];
       properties?: {
         [key: string]: {
           type?: string | string[];
           items?: any;
           description?: string;
           enum?: any[];
         };
       };
     };
@@ -110,6 +112,8 @@
   keep_alive?: string | number // a number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc)
   tools?: Tool[]
   think?: boolean | 'high' | 'medium' | 'low'
+  logprobs?: boolean
+  top_logprobs?: number
 
   options?: Partial<Options>
 }
@@ -174,6 +178,14 @@
 }
 
 // response types
+export interface TokenLogprob {
+  token: string
+  logprob: number
+}
+
+export interface Logprob extends TokenLogprob {
+  top_logprobs?: TokenLogprob[]
+}
 
 export interface GenerateResponse {
   model: string
@@ -189,6 +201,7 @@
   prompt_eval_duration: number
   eval_count: number
   eval_duration: number
+  logprobs?: Logprob[]
 }
 
 export interface ChatResponse {
@@ -203,6 +216,7 @@
   prompt_eval_duration: number
   eval_count: number
   eval_duration: number
+  logprobs?: Logprob[]
 }
 
 export interface EmbedResponse {
@@ -253,9 +267,9 @@
   details: ModelDetails
   messages: Message[]
   modified_at: Date
   model_info: Map<string, any>,
   capabilities: string[],
   projector_info?: Map<string, any>
 }
 
 export interface ListResponse {

test/browser.test.ts

@@ -0,0 +1,49 @@
+import { describe, it, expect, vi } from 'vitest'
+import { Ollama } from '../src/browser'
+import type { ChatResponse, GenerateResponse } from '../src/interfaces'
+
+describe('Ollama logprob request fields', () => {
+  it('forwards logprob settings in generate requests', async () => {
+    const client = new Ollama()
+    const spy = vi
+      .spyOn(client as any, 'processStreamableRequest')
+      .mockResolvedValue({} as GenerateResponse)
+
+    await client.generate({
+      model: 'dummy',
+      prompt: 'Hello',
+      logprobs: true,
+      top_logprobs: 5,
+    })
+
+    expect(spy).toHaveBeenCalledWith(
+      'generate',
+      expect.objectContaining({
+        logprobs: true,
+        top_logprobs: 5,
+      }),
+    )
+  })
+
+  it('forwards logprob settings in chat requests', async () => {
+    const client = new Ollama()
+    const spy = vi
+      .spyOn(client as any, 'processStreamableRequest')
+      .mockResolvedValue({} as ChatResponse)
+
+    await client.chat({
+      model: 'dummy',
+      messages: [{ role: 'user', content: 'hi' }],
+      logprobs: true,
+      top_logprobs: 3,
+    })
+
+    expect(spy).toHaveBeenCalledWith(
+      'chat',
+      expect.objectContaining({
+        logprobs: true,
+        top_logprobs: 3,
+      }),
+    )
+  })
+})