1.7.1: Release (Ollama chat). #270

README.md

@@ -21,7 +21,7 @@ shows the current developments and future ideas.
 - Got a suggestion? [_Add your roadmap ideas_](https://github.com/enricoros/big-agi/issues/new?&template=roadmap-request.md)
 - Want to contribute? [_Pick up a task!_](https://github.com/users/enricoros/projects/4/views/4) - _easy_ to _pro_
 
-### What's New in 1.7.0 · Dec 10, 2023 · Attachment Theory 🌟
+### What's New in 1.7.1 · Dec 11, 2023 · Attachment Theory 🌟
 
 - **Attachments System Overhaul**: Drag, paste, link, snap, text, images, PDFs and more. [#251](https://github.com/enricoros/big-agi/issues/251)
 - **Desktop Webcam Capture**: Image capture now available as Labs feature. [#253](https://github.com/enricoros/big-agi/issues/253)
@@ -31,6 +31,7 @@ shows the current developments and future ideas.
 - Optimized Voice Input and Performance
 - Latest Ollama and Oobabooga models
 - For developers: **Password Protection**: HTTP Basic Auth. [Learn How](https://github.com/enricoros/big-agi/blob/main/docs/deploy-authentication.md)
+- [1.7.1]: Improved Ollama chats. [#270](https://github.com/enricoros/big-agi/issues/270)
 
 ### What's New in 1.6.0 - Nov 28, 2023
 

docs/changelog.md

@@ -10,7 +10,7 @@ by release.
 - work in progress: [big-AGI open roadmap](https://github.com/users/enricoros/projects/4/views/2), [help here](https://github.com/users/enricoros/projects/4/views/4)
 - milestone: [1.8.0](https://github.com/enricoros/big-agi/milestone/8)
 
-### What's New in 1.7.0 · Dec 10, 2023 · Attachment Theory 🌟
+### What's New in 1.7.1 · Dec 11, 2023 · Attachment Theory 🌟
 
 - **Attachments System Overhaul**: Drag, paste, link, snap, text, images, PDFs and more. [#251](https://github.com/enricoros/big-agi/issues/251)
 - **Desktop Webcam Capture**: Image capture now available as Labs feature. [#253](https://github.com/enricoros/big-agi/issues/253)
@@ -20,6 +20,7 @@ by release.
 - Optimized Voice Input and Performance
 - Latest Ollama and Oobabooga models
 - For developers: **Password Protection**: HTTP Basic Auth. [Learn How](https://github.com/enricoros/big-agi/blob/main/docs/deploy-authentication.md)
+- [1.7.1]: Improved Ollama chats. [#270](https://github.com/enricoros/big-agi/issues/270)
 
 ### What's New in 1.6.0 - Nov 28, 2023 · Surf's Up
 

docs/config-ollama.md

@@ -5,15 +5,20 @@ This guide helps you connect [Ollama](https://ollama.ai) [models](https://ollama
 experience. The integration brings the popular big-AGI features to Ollama, including: voice chats,
 editing tools, models switching, personas, and more.
 
+_Last updated Dec 11, 2023_
+
 ![config-local-ollama-0-example.png](pixels/config-ollama-0-example.png)
 
 ## Quick Integration Guide
 
-1. **Ensure Ollama API Server is Running**: Before starting, make sure your Ollama API server is up and running.
-2. **Add Ollama as a Model Source**: In `big-AGI`, navigate to the **Models** section, select **Add a model source**, and choose **Ollama**.
-3. **Enter Ollama Host URL**: Provide the Ollama Host URL where the API server is accessible (e.g., `http://localhost:11434`).
-4. **Refresh Model List**: Once connected, refresh the list of available models to include the Ollama models.
-5. **Start Using AI Personas**: Select an Ollama model and begin interacting with AI personas tailored to your needs.
+1. **Ensure Ollama API Server is Running**: Follow the official instructions to get Ollama up and running on your machine
+2. **Add Ollama as a Model Source**: In `big-AGI`, navigate to the **Models** section, select **Add a model source**, and choose **Ollama**
+3. **Enter Ollama Host URL**: Provide the Ollama Host URL where the API server is accessible (e.g., `http://localhost:11434`)
+4. **Refresh Model List**: Once connected, refresh the list of available models to include the Ollama models
+   > Optional: use the Ollama Admin interface to see which models are available and 'Pull' them in your local machine. Note
+   that this operation will likely timeout due to Edge Functions timeout on the big-AGI server while pulling, and
+   you'll have to press the 'Pull' button again, until a green message appears.
+5. **Chat with Ollama models**: select an Ollama model and begin chatting with AI personas
 
 ### Ollama: installation and Setup
 

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "big-agi",
-  "version": "1.7.0",
+  "version": "1.7.1",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "big-agi",
-      "version": "1.7.0",
+      "version": "1.7.1",
       "hasInstallScript": true,
       "dependencies": {
         "@dqbd/tiktoken": "^1.0.7",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "big-agi",
-  "version": "1.7.0",
+  "version": "1.7.1",
   "private": true,
   "scripts": {
     "dev": "next dev",

src/apps/news/news.data.tsx

@@ -67,9 +67,10 @@ export const NewsItems: NewsItem[] = [
     ],
   },*/
   {
-    versionCode: '1.7.0',
+    versionCode: '1.7.1',
     versionName: 'Attachment Theory',
-    versionDate: new Date('2023-12-10T12:00:00Z'), // new Date().toISOString()
+    versionDate: new Date('2023-12-11T06:00:00Z'), // new Date().toISOString()
+    // versionDate: new Date('2023-12-10T12:00:00Z'), // 1.7.0
     items: [
       { text: <>Redesigned <B href={RIssues + '/251'}>attachments system</B>: drag, paste, link, snap, images, text, pdfs</> },
       { text: <>Desktop <B href={RIssues + '/253'}>webcam access</B> for direct image capture (Labs option)</> },
@@ -79,6 +80,7 @@ export const NewsItems: NewsItem[] = [
       { text: <>{platformAwareKeystrokes('Ctrl+Shift+O')}: quick access to model options</> },
       { text: <>Optimized voice input and performance</> },
       { text: <>Latest Ollama and Oobabooga models</> },
+      { text: <>1.7.1: Improved <B href={RIssues + '/270'}>Ollama chats</B></> },
     ],
   },
   {

src/common/app.theme.ts

@@ -46,6 +46,7 @@ export const appTheme = extendTheme({
         text: {
           icon: 'var(--joy-palette-neutral-700)',           // <IconButton color='neutral' /> icon color
           secondary: 'var(--joy-palette-neutral-800)',      // increase contrast a bit
+          // tertiary: 'var(--joy-palette-neutral-700)',       // increase contrast a bit
         },
         // popup [white] > surface [50] > level1 [100] > level2 [200] > level3 [300] > body [white -> 400]
         background: {

src/modules/llms/transports/server/ollama/ollama.models.ts

@@ -3,54 +3,57 @@
  * descriptions for the models.
  * (nor does it reliably provide context window sizes) - TODO: open a bug upstream
  *
- * from: https://ollama.ai/library?sort=popular
+ * from: https://ollama.ai/library?sort=featured
  */
 export const OLLAMA_BASE_MODELS: { [key: string]: { description: string, pulls: number, added?: string } } = {
-  'mistral': { description: 'The Mistral 7B model released by Mistral AI', pulls: 56100 },
-  'llama2': { description: 'The most popular model for general use.', pulls: 117400 },
-  'codellama': { description: 'A large language model that can use text prompts to generate and discuss code.', pulls: 61500 },
-  'llama2-uncensored': { description: 'Uncensored Llama 2 model by George Sung and Jarrad Hope.', pulls: 26800 },
-  'orca-mini': { description: 'A general-purpose model ranging from 3 billion parameters to 70 billion, suitable for entry-level hardware.', pulls: 23000 },
-  'vicuna': { description: 'General use chat model based on Llama and Llama 2 with 2K to 16K context sizes.', pulls: 20600 },
-  'wizard-vicuna-uncensored': { description: 'Wizard Vicuna Uncensored is a 7B, 13B, and 30B parameter model based on Llama 2 uncensored by Eric Hartford.', pulls: 12100 },
-  'phind-codellama': { description: 'Code generation model based on CodeLlama.', pulls: 9760 },
-  'wizardcoder': { description: 'Llama based code generation model focused on Python.', pulls: 9002 },
-  'mistral-openorca': { description: 'Mistral OpenOrca is a 7 billion parameter model, fine-tuned on top of the Mistral 7B model using the OpenOrca dataset.', pulls: 8671 },
-  'nous-hermes': { description: 'General use models based on Llama and Llama 2 from Nous Research.', pulls: 8478 },
-  'zephyr': { description: 'Zephyr beta is a fine-tuned 7B version of mistral that was trained on on a mix of publicly available, synthetic datasets.', pulls: 8142 },
-  'wizard-math': { description: 'Model focused on math and logic problems', pulls: 7426 },
-  'llama2-chinese': { description: 'Llama 2 based model fine tuned to improve Chinese dialogue ability.', pulls: 7035 },
-  'stable-beluga': { description: 'Llama 2 based model fine tuned on an Orca-style dataset. Originally called Free Willy.', pulls: 6140 },
-  'falcon': { description: 'A large language model built by the Technology Innovation Institute (TII) for use in summarization, text generation, and chat bots.', pulls: 5865 },
-  'codeup': { description: 'Great code generation model based on Llama2.', pulls: 5534 },
-  'everythinglm': { description: 'Uncensored Llama2 based model with 16k context size.', pulls: 4696 },
-  'medllama2': { description: 'Fine-tuned Llama 2 model to answer medical questions based on an open source medical dataset.', pulls: 4275 },
-  'wizardlm-uncensored': { description: 'Uncensored version of Wizard LM model.', pulls: 4227 },
-  'deepseek-coder': { description: 'DeepSeek Coder is trained from scratch on both 87% code and 13% natural language in English and Chinese. Each of the models are pre-trained on 2 trillion tokens.', pulls: 3663, added: '20231129' },
-  'wizard-vicuna': { description: 'Wizard Vicuna is a 13B parameter model based on Llama 2 trained by MelodysDreamj.', pulls: 3343 },
-  'orca2': { description: 'Orca 2 is built by Microsoft research, and are a fine-tuned version of Meta\'s Llama 2 models. The model is designed to excel particularly in reasoning.', pulls: 3134, added: '20231129' },
-  'open-orca-platypus2': { description: 'Merge of the Open Orca OpenChat model and the Garage-bAInd Platypus 2 model. Designed for chat and code generation.', pulls: 3050 },
-  'starcoder': { description: 'StarCoder is a code generation model trained on 80+ programming languages.', pulls: 2981 },
-  'dolphin2.2-mistral': { description: 'An instruct-tuned model based on Mistral. Version 2.2 is fine-tuned for improved conversation and empathy.', pulls: 2636 },
-  'yarn-mistral': { description: 'An extension of Mistral to support a context of up to 128k tokens.', pulls: 2328 },
-  'openchat': { description: 'A family of open-source models trained on a wide variety of data, surpassing ChatGPT on various benchmarks.', pulls: 2281, added: '20231129' },
-  'openhermes2.5-mistral': { description: 'OpenHermes 2.5 Mistral 7B is a Mistral 7B fine-tune, a continuation of OpenHermes 2 model, which trained on additional code datasets.', pulls: 2101 },
-  'yi': { description: 'A high-performing, bilingual base model.', pulls: 1806 },
-  'samantha-mistral': { description: 'A companion assistant trained in philosophy, psychology, and personal relationships. Based on Mistral.', pulls: 1803 },
-  'yarn-llama2': { description: 'An extension of Llama 2 that supports a context of up to 128k tokens.', pulls: 1605 },
-  'sqlcoder': { description: 'SQLCoder is a code completion model fined-tuned on StarCoder for SQL generation tasks.', pulls: 1584 },
-  'openhermes2-mistral': { description: 'OpenHermes 2 Mistral is a 7B model fine-tuned on Mistral with 900,000 entries of primarily GPT-4 generated data from open datasets.', pulls: 1560 },
-  'neural-chat': { description: 'A fine-tuned model based on Mistral with good coverage of domain and language.', pulls: 1338, added: '20231129' },
-  'wizardlm': { description: 'General use 70 billion parameter model based on Llama 2.', pulls: 1253 },
-  'dolphin2.1-mistral': { description: 'An instruct-tuned model based on Mistral and trained on a dataset filtered to remove alignment and bias.', pulls: 1163 },
-  'mistrallite': { description: 'MistralLite is a fine-tuned model based on Mistral with enhanced capabilities of processing long contexts.', pulls: 1099 },
-  'codebooga': { description: 'A high-performing code instruct model created by merging two existing code models.', pulls: 1042 },
-  'goliath': { description: 'A language model created by combining two fine-tuned Llama 2 70B models into one.', pulls: 728, added: '20231129' },
-  'xwinlm': { description: 'Conversational model based on Llama 2 that performs competitively on various benchmarks.', pulls: 593 },
-  'nexusraven': { description: 'Nexus Raven is a 13B instruction tuned model for function calling tasks.', pulls: 585 },
-  'alfred': { description: 'A robust conversational model designed to be used for both chat and instruct use cases.', pulls: 573, added: '20231129' },
-  'starling-lm': { description: 'Starling is a large language model trained by reinforcement learning from AI feedback focused on improving chatbot helpfulness.', pulls: 446, added: '20231129' },
-  'meditron': { description: 'Open-source medical large language model adapted from Llama 2 to the medical domain.', pulls: 100, added: '20231129' },
-  'deepseek-llm': { description: 'An advanced language model crafted with 2 trillion bilingual tokens.', pulls: 11, added: '20231129' },
+  'starling-lm': { description: 'Starling is a large language model trained by reinforcement learning from AI feedback focused on improving chatbot helpfulness.', pulls: 2353, added: '20231129' },
+  'neural-chat': { description: 'A fine-tuned model based on Mistral with good coverage of domain and language.', pulls: 3089, added: '20231129' },
+  'mistral': { description: 'The Mistral 7B model released by Mistral AI', pulls: 70300 },
+  'yi': { description: 'A high-performing, bilingual base model.', pulls: 2673 },
+  'llama2': { description: 'The most popular model for general use.', pulls: 141000 },
+  'codellama': { description: 'A large language model that can use text prompts to generate and discuss code.', pulls: 71400 },
+  'llama2-uncensored': { description: 'Uncensored Llama 2 model by George Sung and Jarrad Hope.', pulls: 30900 },
+  'orca-mini': { description: 'A general-purpose model ranging from 3 billion parameters to 70 billion, suitable for entry-level hardware.', pulls: 26000 },
+  'vicuna': { description: 'General use chat model based on Llama and Llama 2 with 2K to 16K context sizes.', pulls: 21800 },
+  'wizard-vicuna-uncensored': { description: 'Wizard Vicuna Uncensored is a 7B, 13B, and 30B parameter model based on Llama 2 uncensored by Eric Hartford.', pulls: 13700 },
+  'phind-codellama': { description: 'Code generation model based on CodeLlama.', pulls: 10600 },
+  'zephyr': { description: 'Zephyr beta is a fine-tuned 7B version of mistral that was trained on on a mix of publicly available, synthetic datasets.', pulls: 10200 },
+  'wizardcoder': { description: 'Llama based code generation model focused on Python.', pulls: 9895 },
+  'mistral-openorca': { description: 'Mistral OpenOrca is a 7 billion parameter model, fine-tuned on top of the Mistral 7B model using the OpenOrca dataset.', pulls: 9256 },
+  'nous-hermes': { description: 'General use models based on Llama and Llama 2 from Nous Research.', pulls: 8827 },
+  'wizard-math': { description: 'Model focused on math and logic problems', pulls: 7849 },
+  'llama2-chinese': { description: 'Llama 2 based model fine tuned to improve Chinese dialogue ability.', pulls: 7375 },
+  'deepseek-coder': { description: 'DeepSeek Coder is trained from scratch on both 87% code and 13% natural language in English and Chinese. Each of the models are pre-trained on 2 trillion tokens.', pulls: 7335, added: '20231129' },
+  'falcon': { description: 'A large language model built by the Technology Innovation Institute (TII) for use in summarization, text generation, and chat bots.', pulls: 6726 },
+  'stable-beluga': { description: 'Llama 2 based model fine tuned on an Orca-style dataset. Originally called Free Willy.', pulls: 6272 },
+  'codeup': { description: 'Great code generation model based on Llama2.', pulls: 5978 },
+  'orca2': { description: 'Orca 2 is built by Microsoft research, and are a fine-tuned version of Meta\'s Llama 2 models. The model is designed to excel particularly in reasoning.', pulls: 5854, added: '20231129' },
+  'everythinglm': { description: 'Uncensored Llama2 based model with 16k context size.', pulls: 5040 },
+  'medllama2': { description: 'Fine-tuned Llama 2 model to answer medical questions based on an open source medical dataset.', pulls: 4648 },
+  'wizardlm-uncensored': { description: 'Uncensored version of Wizard LM model.', pulls: 4536 },
+  'dolphin2.2-mistral': { description: 'An instruct-tuned model based on Mistral. Version 2.2 is fine-tuned for improved conversation and empathy.', pulls: 3638 },
+  'starcoder': { description: 'StarCoder is a code generation model trained on 80+ programming languages.', pulls: 3638 },
+  'wizard-vicuna': { description: 'Wizard Vicuna is a 13B parameter model based on Llama 2 trained by MelodysDreamj.', pulls: 3485 },
+  'openchat': { description: 'A family of open-source models trained on a wide variety of data, surpassing ChatGPT on various benchmarks.', pulls: 3438, added: '20231129' },
+  'open-orca-platypus2': { description: 'Merge of the Open Orca OpenChat model and the Garage-bAInd Platypus 2 model. Designed for chat and code generation.', pulls: 3145 },
+  'openhermes2.5-mistral': { description: 'OpenHermes 2.5 Mistral 7B is a Mistral 7B fine-tune, a continuation of OpenHermes 2 model, which trained on additional code datasets.', pulls: 3023 },
+  'yarn-mistral': { description: 'An extension of Mistral to support a context of up to 128k tokens.', pulls: 2775 },
+  'samantha-mistral': { description: 'A companion assistant trained in philosophy, psychology, and personal relationships. Based on Mistral.', pulls: 2192 },
+  'sqlcoder': { description: 'SQLCoder is a code completion model fined-tuned on StarCoder for SQL generation tasks', pulls: 1973 },
+  'yarn-llama2': { description: 'An extension of Llama 2 that supports a context of up to 128k tokens.', pulls: 1915 },
+  'openhermes2-mistral': { description: 'OpenHermes 2 Mistral is a 7B model fine-tuned on Mistral with 900,000 entries of primarily GPT-4 generated data from open datasets.', pulls: 1690 },
+  'meditron': { description: 'Open-source medical large language model adapted from Llama 2 to the medical domain.', pulls: 1667, added: '20231129' },
+  'wizardlm': { description: 'General use 70 billion parameter model based on Llama 2.', pulls: 1379 },
+  'mistrallite': { description: 'MistralLite is a fine-tuned model based on Mistral with enhanced capabilities of processing long contexts.', pulls: 1345 },
+  'deepseek-llm': { description: 'An advanced language model crafted with 2 trillion bilingual tokens.', pulls: 1318, added: '20231129' },
+  'dolphin2.1-mistral': { description: 'An instruct-tuned model based on Mistral and trained on a dataset filtered to remove alignment and bias.', pulls: 1302 },
+  'codebooga': { description: 'A high-performing code instruct model created by merging two existing code models.', pulls: 1254 },
+  'goliath': { description: 'A language model created by combining two fine-tuned Llama 2 70B models into one.', pulls: 946, added: '20231129' },
+  'stablelm-zephyr': { description: 'A lightweight chat model allowing accurate, and responsive output without requiring high-end hardware.', pulls: 945, added: '20231210' },
+  'nexusraven': { description: 'Nexus Raven is a 13B instruction tuned model for function calling tasks.', pulls: 860 },
+  'magicoder': { description: '🎩 Magicoder is a family of 7B parameter models trained on 75K synthetic instruction data using OSS-Instruct, a novel approach to enlightening LLMs with open-source code snippets.', pulls: 816, added: '20231210' },
+  'alfred': { description: 'A robust conversational model designed to be used for both chat and instruct use cases.', pulls: 804, added: '20231129' },
+  'xwinlm': { description: 'Conversational model based on Llama 2 that performs competitively on various benchmarks.', pulls: 706 },
 };
-export const OLLAMA_LAST_UPDATE: string = '20231129';
+// export const OLLAMA_LAST_UPDATE: string = '20231210';
+export const OLLAMA_PREV_UPDATE: string = '20231129';

src/modules/llms/transports/server/ollama/ollama.router.ts

@@ -11,12 +11,15 @@ import { capitalizeFirstLetter } from '~/common/util/textUtils';
 import { fixupHost, openAIChatGenerateOutputSchema, OpenAIHistorySchema, openAIHistorySchema, OpenAIModelSchema, openAIModelSchema } from '../openai/openai.router';
 import { listModelsOutputSchema, ModelDescriptionSchema } from '../server.schemas';
 
-import { OLLAMA_BASE_MODELS, OLLAMA_LAST_UPDATE } from './ollama.models';
-import { wireOllamaGenerationSchema } from './ollama.wiretypes';
+import { OLLAMA_BASE_MODELS, OLLAMA_PREV_UPDATE } from './ollama.models';
+import { WireOllamaChatCompletionInput, wireOllamaChunkedOutputSchema } from './ollama.wiretypes';
 
 
 // Default hosts
 const DEFAULT_OLLAMA_HOST = 'http://127.0.0.1:11434';
+export const OLLAMA_PATH_CHAT = '/api/chat';
+const OLLAMA_PATH_TAGS = '/api/tags';
+const OLLAMA_PATH_SHOW = '/api/show';
 
 
 // Mappers
@@ -34,7 +37,23 @@ export function ollamaAccess(access: OllamaAccessSchema, apiPath: string): { hea
 
 }
 
-export function ollamaChatCompletionPayload(model: OpenAIModelSchema, history: OpenAIHistorySchema, stream: boolean) {
+
+export const ollamaChatCompletionPayload = (model: OpenAIModelSchema, history: OpenAIHistorySchema, stream: boolean): WireOllamaChatCompletionInput => ({
+  model: model.id,
+  messages: history,
+  options: {
+    ...(model.temperature && { temperature: model.temperature }),
+  },
+  // n: ...
+  // functions: ...
+  // function_call: ...
+  stream,
+});
+
+
+/* Unused: switched to the Chat endpoint (above). The implementation is left here for reference.
+https://github.com/jmorganca/ollama/blob/main/docs/api.md#generate-a-completion
+export function ollamaCompletionPayload(model: OpenAIModelSchema, history: OpenAIHistorySchema, stream: boolean) {
 
   // if the first message is the system prompt, extract it
   let systemPrompt: string | undefined = undefined;
@@ -62,7 +81,7 @@ export function ollamaChatCompletionPayload(model: OpenAIModelSchema, history: O
     ...(systemPrompt && { system: systemPrompt }),
     stream,
   };
-}
+}*/
 
 async function ollamaGET<TOut extends object>(access: OllamaAccessSchema, apiPath: string /*, signal?: AbortSignal*/): Promise<TOut> {
   const { headers, url } = ollamaAccess(access, apiPath);
@@ -104,6 +123,7 @@ const listPullableOutputSchema = z.object({
     label: z.string(),
     tag: z.string(),
     description: z.string(),
+    pulls: z.number(),
     isNew: z.boolean(),
   })),
 });
@@ -122,7 +142,8 @@ export const llmOllamaRouter = createTRPCRouter({
           label: capitalizeFirstLetter(model_id),
           tag: 'latest',
           description: model.description,
-          isNew: !!model.added && model.added >= OLLAMA_LAST_UPDATE,
+          pulls: model.pulls,
+          isNew: !!model.added && model.added >= OLLAMA_PREV_UPDATE,
         })),
       };
     }),
@@ -160,6 +181,7 @@ export const llmOllamaRouter = createTRPCRouter({
         throw new Error('Ollama delete issue: ' + deleteOutput);
     }),
 
+
   /* Ollama: List the Models available */
   listModels: publicProcedure
     .input(accessOnlySchema)
@@ -167,7 +189,7 @@ export const llmOllamaRouter = createTRPCRouter({
     .query(async ({ input }) => {
 
       // get the models
-      const wireModels = await ollamaGET(input.access, '/api/tags');
+      const wireModels = await ollamaGET(input.access, OLLAMA_PATH_TAGS);
       const wireOllamaListModelsSchema = z.object({
         models: z.array(z.object({
           name: z.string(),
@@ -180,7 +202,7 @@ export const llmOllamaRouter = createTRPCRouter({
 
       // retrieve info for each of the models (/api/show, post call, in parallel)
       const detailedModels = await Promise.all(models.map(async model => {
-        const wireModelInfo = await ollamaPOST(input.access, { 'name': model.name }, '/api/show');
+        const wireModelInfo = await ollamaPOST(input.access, { 'name': model.name }, OLLAMA_PATH_SHOW);
         const wireOllamaModelInfoSchema = z.object({
           license: z.string().optional(),
           modelfile: z.string(),
@@ -221,12 +243,15 @@ export const llmOllamaRouter = createTRPCRouter({
     .output(openAIChatGenerateOutputSchema)
     .mutation(async ({ input: { access, history, model } }) => {
 
-      const wireGeneration = await ollamaPOST(access, ollamaChatCompletionPayload(model, history, false), '/api/generate');
-      const generation = wireOllamaGenerationSchema.parse(wireGeneration);
+      const wireGeneration = await ollamaPOST(access, ollamaChatCompletionPayload(model, history, false), OLLAMA_PATH_CHAT);
+      const generation = wireOllamaChunkedOutputSchema.parse(wireGeneration);
+
+      if (!generation.message?.content)
+        throw new Error('Ollama chat generation (non-stream) issue: ' + JSON.stringify(wireGeneration));
 
       return {
         role: 'assistant',
-        content: generation.response,
+        content: generation.message.content,
         finish_reason: generation.done ? 'stop' : null,
       };
     }),

src/modules/llms/transports/server/ollama/ollama.wiretypes.ts

@@ -1,16 +1,69 @@
 import { z } from 'zod';
 
-export const wireOllamaGenerationSchema = z.object({
+
+/**
+ * Chat Completion API - Request
+ * https://github.com/jmorganca/ollama/blob/main/docs/api.md#generate-a-chat-completion
+ */
+const wireOllamaChatCompletionInputSchema = z.object({
+
+  // required
+  model: z.string(),
+  messages: z.array(z.object({
+    role: z.enum(['assistant', 'system', 'user']),
+    content: z.string(),
+  })),
+
+  // optional
+  format: z.enum(['json']).optional(),
+  options: z.object({
+    // https://github.com/jmorganca/ollama/blob/main/docs/modelfile.md
+    // Maximum number of tokens to predict when generating text.
+    num_predict: z.number().int().optional(),
+    // Sets the random number seed to use for generation
+    seed: z.number().int().optional(),
+    // The temperature of the model
+    temperature: z.number().positive().optional(),
+    // Reduces the probability of generating nonsense (Default: 40)
+    top_k: z.number().positive().optional(),
+    // Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text. (Default 0.9)
+    top_p: z.number().positive().optional(),
+  }).optional(),
+  template: z.string().optional(), // overrides what is defined in the Modelfile
+  stream: z.boolean().optional(), // default: true
+
+  // Future Improvements?
+  // n: z.number().int().optional(), // number of completions to generate
+  // functions: ...
+  // function_call: ...
+});
+export type WireOllamaChatCompletionInput = z.infer<typeof wireOllamaChatCompletionInputSchema>;
+
+
+/**
+ * Chat Completion or Generation APIs - Streaming Response
+ */
+export const wireOllamaChunkedOutputSchema = z.object({
   model: z.string(),
   // created_at: z.string(), // commented because unused
-  response: z.string(),
+
+  // [Chat Completion] (exclusive with 'response')
+  message: z.object({
+    role: z.enum(['assistant' /*, 'system', 'user' Disabled on purpose, to validate the response */]),
+    content: z.string(),
+  }).optional(), // optional on the last message
+
+  // [Generation] (non-chat, exclusive with 'message')
+  //response: z.string().optional(),
+
   done: z.boolean(),
 
   // only on the last message
-  // context: z.array(z.number()),
+  // context: z.array(z.number()), // non-chat endpoint
   // total_duration: z.number(),
-  // load_duration: z.number(),
-  // eval_duration: z.number(),
   // prompt_eval_count: z.number(),
+  // prompt_eval_duration: z.number(),
   // eval_count: z.number(),
-});
+  // eval_duration: z.number(),
+
+});

src/modules/llms/transports/server/openai/openai.streaming.ts

@@ -6,10 +6,10 @@ import { createEmptyReadableStream, debugGenerateCurlCommand, safeErrorString, S
 
 import type { AnthropicWire } from '../anthropic/anthropic.wiretypes';
 import type { OpenAIWire } from './openai.wiretypes';
+import { OLLAMA_PATH_CHAT, ollamaAccess, ollamaAccessSchema, ollamaChatCompletionPayload } from '../ollama/ollama.router';
 import { anthropicAccess, anthropicAccessSchema, anthropicChatCompletionPayload } from '../anthropic/anthropic.router';
-import { ollamaAccess, ollamaAccessSchema, ollamaChatCompletionPayload } from '../ollama/ollama.router';
 import { openAIAccess, openAIAccessSchema, openAIChatCompletionPayload, openAIHistorySchema, openAIModelSchema } from './openai.router';
-import { wireOllamaGenerationSchema } from '../ollama/ollama.wiretypes';
+import { wireOllamaChunkedOutputSchema } from '../ollama/ollama.wiretypes';
 
 
 /**
@@ -59,10 +59,10 @@ export async function openaiStreamingRelayHandler(req: NextRequest): Promise<Res
         break;
 
       case 'ollama':
-        headersUrl = ollamaAccess(access, '/api/generate');
+        headersUrl = ollamaAccess(access, OLLAMA_PATH_CHAT);
         body = ollamaChatCompletionPayload(model, history, true);
         eventStreamFormat = 'json-nl';
-        vendorStreamParser = createOllamaStreamParser();
+        vendorStreamParser = createOllamaChatCompletionStreamParser();
         break;
 
       case 'azure':
@@ -135,30 +135,35 @@ function createAnthropicStreamParser(): AIStreamParser {
   };
 }
 
-function createOllamaStreamParser(): AIStreamParser {
+function createOllamaChatCompletionStreamParser(): AIStreamParser {
   let hasBegun = false;
 
   return (data: string) => {
 
-    let wireGeneration: any;
+    // parse the JSON chunk
+    let wireJsonChunk: any;
     try {
-      wireGeneration = JSON.parse(data);
+      wireJsonChunk = JSON.parse(data);
     } catch (error: any) {
       // log the malformed data to the console, and rethrow to transmit as 'error'
       console.log(`/api/llms/stream: Ollama parsing issue: ${error?.message || error}`, data);
       throw error;
     }
-    const generation = wireOllamaGenerationSchema.parse(wireGeneration);
-    let text = generation.response;
+
+    // validate chunk
+    const chunk = wireOllamaChunkedOutputSchema.parse(wireJsonChunk);
+
+    // process output
+    let text = chunk.message?.content || /*chunk.response ||*/ '';
 
     // hack: prepend the model name to the first packet
-    if (!hasBegun) {
+    if (!hasBegun && chunk.model) {
       hasBegun = true;
-      const firstPacket: ChatStreamFirstPacketSchema = { model: generation.model };
+      const firstPacket: ChatStreamFirstPacketSchema = { model: chunk.model };
       text = JSON.stringify(firstPacket) + text;
     }
 
-    return { text, close: generation.done };
+    return { text, close: chunk.done };
   };
 }
 
@@ -248,7 +253,8 @@ function createEventStreamTransformer(vendorTextParser: AIStreamParser, inputFor
           if (close)
             controller.terminate();
         } catch (error: any) {
-          // console.log(`/api/llms/stream: parse issue: ${error?.message || error}`);
+          if (SERVER_DEBUG_WIRE)
+            console.log(' - E: parse issue:', event.data, error?.message || error);
           controller.enqueue(textEncoder.encode(`[Stream Issue] ${dialectLabel}: ${safeErrorString(error) || 'Unknown stream parsing error'}`));
           controller.terminate();
         }

src/modules/llms/vendors/ollama/OllamaAdministration.tsx

@@ -1,24 +1,29 @@
 import * as React from 'react';
 
-import { Box, Button, Chip, FormControl, Input, Option, Select, Stack, Typography } from '@mui/joy';
+import { Box, Button, Chip, FormControl, IconButton, Input, Option, Select, Stack, Typography } from '@mui/joy';
+import LaunchIcon from '@mui/icons-material/Launch';
+import FormatListNumberedRtlIcon from '@mui/icons-material/FormatListNumberedRtl';
 
 import { FormLabelStart } from '~/common/components/forms/FormLabelStart';
 import { GoodModal } from '~/common/components/GoodModal';
+import { GoodTooltip } from '~/common/components/GoodTooltip';
+import { InlineError } from '~/common/components/InlineError';
+import { Link } from '~/common/components/Link';
 import { apiQuery } from '~/common/util/trpc.client';
 import { settingsGap } from '~/common/app.theme';
 
 import type { OllamaAccessSchema } from '../../transports/server/ollama/ollama.router';
-import { InlineError } from '~/common/components/InlineError';
 
 
-export function OllamaAdmin(props: { access: OllamaAccessSchema, onClose: () => void }) {
+export function OllamaAdministration(props: { access: OllamaAccessSchema, onClose: () => void }) {
 
   // state
+  const [sortByPulls, setSortByPulls] = React.useState<boolean>(false);
   const [modelName, setModelName] = React.useState<string | null>('llama2');
   const [modelTag, setModelTag] = React.useState<string>('');
 
   // external state
-  const { data: pullable } = apiQuery.llmOllama.adminListPullable.useQuery({ access: props.access }, {
+  const { data: pullableData } = apiQuery.llmOllama.adminListPullable.useQuery({ access: props.access }, {
     staleTime: 1000 * 60,
     refetchOnWindowFocus: false,
   });
@@ -26,7 +31,11 @@ export function OllamaAdmin(props: { access: OllamaAccessSchema, onClose: () =>
   const { isLoading: isDeleting, status: deleteStatus, error: deleteError, mutate: deleteMutate, reset: deleteReset } = apiQuery.llmOllama.adminDelete.useMutation();
 
   // derived state
-  const pullModelDescription = pullable?.pullable.find(p => p.id === modelName)?.description ?? null;
+  let pullable = pullableData?.pullable || [];
+  if (sortByPulls)
+    pullable = pullable.toSorted((a, b) => b.pulls - a.pulls);
+  const pullModelDescription = pullable.find(p => p.id === modelName)?.description ?? null;
+
 
   const handleModelPull = () => {
     deleteReset();
@@ -38,6 +47,7 @@ export function OllamaAdmin(props: { access: OllamaAccessSchema, onClose: () =>
     modelName && deleteMutate({ access: props.access, name: modelName + (modelTag ? ':' + modelTag : '') });
   };
 
+
   return (
     <GoodModal title='Ollama Administration' dividers open onClose={props.onClose}>
 
@@ -47,25 +57,48 @@ export function OllamaAdmin(props: { access: OllamaAccessSchema, onClose: () =>
           However we provide a way to pull models from the Ollama host, for convenience.
         </Typography>
 
-        <Box sx={{ display: 'flex', gap: 1 }}>
-          <FormControl sx={{ flexGrow: 1 }}>
+        <Box sx={{ display: 'flex', flexFlow: 'row wrap', gap: 1 }}>
+          <FormControl sx={{ flexGrow: 1, flexBasis: 0.55 }}>
             <FormLabelStart title='Name' />
-            <Select value={modelName || ''} onChange={(_event: any, value: string | null) => setModelName(value)}>
-              {pullable?.pullable.map(p =>
-                <Option key={p.id} value={p.id}>
-                  {p.isNew === true && <Chip size='sm' variant='outlined'>New</Chip>} {p.label}
-                </Option>,
-              )}
-            </Select>
+            <Box sx={{ display: 'flex', gap: 1 }}>
+              <Select
+                value={modelName || ''}
+                onChange={(_event: any, value: string | null) => setModelName(value)}
+                sx={{ flexGrow: 1 }}
+              >
+                {pullable.map(p =>
+                  <Option key={p.id} value={p.id}>
+                    {p.isNew === true && <Chip size='sm' variant='outlined'>NEW</Chip>} {p.label}{sortByPulls && ` (${p.pulls.toLocaleString()})`}
+                  </Option>,
+                )}
+              </Select>
+              <GoodTooltip title='Sort by Downloads'>
+                <IconButton
+                  variant={sortByPulls ? 'solid' : 'outlined'}
+                  onClick={() => setSortByPulls(!sortByPulls)}
+                >
+                  <FormatListNumberedRtlIcon />
+                </IconButton>
+              </GoodTooltip>
+            </Box>
           </FormControl>
-          <FormControl sx={{ flexGrow: 1 }}>
+          <FormControl sx={{ flexGrow: 1, flexBasis: 0.45 }}>
             <FormLabelStart title='Tag' />
-            <Input
-              variant='outlined' placeholder='latest'
-              value={modelTag || ''} onChange={event => setModelTag(event.target.value)}
-              sx={{ minWidth: 100 }}
-              slotProps={{ input: { size: 10 } }} // halve the min width
-            />
+            <Box sx={{ display: 'flex', gap: 1 }}>
+              <Input
+                variant='outlined' placeholder='latest'
+                value={modelTag || ''} onChange={event => setModelTag(event.target.value)}
+                sx={{ minWidth: 80, flexGrow: 1 }}
+                slotProps={{ input: { size: 10 } }} // halve the min width
+              />
+              {!!modelName && (
+                <IconButton
+                  component={Link} href={`https://ollama.ai/library/${modelName}`} target='_blank'
+                >
+                  <LaunchIcon />
+                </IconButton>
+              )}
+            </Box>
           </FormControl>
         </Box>
 
@@ -85,7 +118,7 @@ export function OllamaAdmin(props: { access: OllamaAccessSchema, onClose: () =>
             {pullModelDescription}
           </Typography>
 
-          <Box sx={{ display: 'flex', gap: 1 }}>
+          <Box sx={{ display: 'flex', flexWrap: 1, gap: 1 }}>
             <Button
               variant='outlined'
               color={deleteStatus === 'error' ? 'danger' : deleteStatus === 'success' ? 'success' : 'primary'}

src/modules/llms/vendors/ollama/OllamaSourceSetup.tsx

@@ -11,7 +11,7 @@ import { asValidURL } from '~/common/util/urlUtils';
 
 import { DModelSourceId, useModelsStore, useSourceSetup } from '../../store-llms';
 import { ModelVendorOllama } from './ollama.vendor';
-import { OllamaAdmin } from './OllamaAdmin';
+import { OllamaAdministration } from './OllamaAdministration';
 import { modelDescriptionToDLLM } from '../openai/OpenAISourceSetup';
 
 
@@ -63,7 +63,7 @@ export function OllamaSourceSetup(props: { sourceId: DModelSourceId }) {
 
     {isError && <InlineError error={error} />}
 
-    {adminOpen && <OllamaAdmin access={access} onClose={() => setAdminOpen(false)} />}
+    {adminOpen && <OllamaAdministration access={access} onClose={() => setAdminOpen(false)} />}
 
   </>;
 }