1use anyhow::{Context as _, Result};
2use futures::{AsyncBufReadExt, AsyncReadExt, StreamExt, io::BufReader, stream::BoxStream};
3use http_client::{AsyncBody, HttpClient, Method, Request as HttpRequest, http};
4use serde::{Deserialize, Serialize};
5use serde_json::Value;
6pub use settings::KeepAlive;
7use std::time::Duration;
8
9pub const OLLAMA_API_URL: &str = "http://localhost:11434";
10
11#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
12#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
13pub struct Model {
14 pub name: String,
15 pub display_name: Option<String>,
16 pub max_tokens: u64,
17 pub keep_alive: Option<KeepAlive>,
18 pub supports_tools: Option<bool>,
19 pub supports_vision: Option<bool>,
20 pub supports_thinking: Option<bool>,
21}
22
23fn get_max_tokens(name: &str) -> u64 {
24 /// Default context length for unknown models.
25 const DEFAULT_TOKENS: u64 = 4096;
26 /// Magic number. Lets many Ollama models work with ~16GB of ram.
27 /// Models that support context beyond 16k such as codestral (32k) or devstral (128k) will be clamped down to 16k
28 const MAXIMUM_TOKENS: u64 = 16384;
29
30 match name.split(':').next().unwrap() {
31 "granite-code" | "phi" | "tinyllama" => 2048,
32 "llama2" | "stablelm2" | "vicuna" | "yi" => 4096,
33 "aya" | "codegemma" | "gemma" | "gemma2" | "llama3" | "starcoder" => 8192,
34 "codellama" | "starcoder2" => 16384,
35 "codestral" | "dolphin-mixtral" | "llava" | "magistral" | "mistral" | "mixstral"
36 | "qwen2" | "qwen2.5-coder" => 32768,
37 "cogito" | "command-r" | "deepseek-coder-v2" | "deepseek-r1" | "deepseek-v3"
38 | "devstral" | "gemma3" | "gpt-oss" | "granite3.3" | "llama3.1" | "llama3.2"
39 | "llama3.3" | "mistral-nemo" | "phi3" | "phi3.5" | "phi4" | "qwen3" | "yi-coder" => 128000,
40 _ => DEFAULT_TOKENS,
41 }
42 .clamp(1, MAXIMUM_TOKENS)
43}
44
45impl Model {
46 pub fn new(
47 name: &str,
48 display_name: Option<&str>,
49 max_tokens: Option<u64>,
50 supports_tools: Option<bool>,
51 supports_vision: Option<bool>,
52 supports_thinking: Option<bool>,
53 ) -> Self {
54 Self {
55 name: name.to_owned(),
56 display_name: display_name
57 .map(ToString::to_string)
58 .or_else(|| name.strip_suffix(":latest").map(ToString::to_string)),
59 max_tokens: max_tokens.unwrap_or_else(|| get_max_tokens(name)),
60 keep_alive: Some(KeepAlive::indefinite()),
61 supports_tools,
62 supports_vision,
63 supports_thinking,
64 }
65 }
66
67 pub fn id(&self) -> &str {
68 &self.name
69 }
70
71 pub fn display_name(&self) -> &str {
72 self.display_name.as_ref().unwrap_or(&self.name)
73 }
74
75 pub fn max_token_count(&self) -> u64 {
76 self.max_tokens
77 }
78}
79
80#[derive(Serialize, Deserialize, Debug)]
81#[serde(tag = "role", rename_all = "lowercase")]
82pub enum ChatMessage {
83 Assistant {
84 content: String,
85 tool_calls: Option<Vec<OllamaToolCall>>,
86 #[serde(skip_serializing_if = "Option::is_none")]
87 images: Option<Vec<String>>,
88 thinking: Option<String>,
89 },
90 User {
91 content: String,
92 #[serde(skip_serializing_if = "Option::is_none")]
93 images: Option<Vec<String>>,
94 },
95 System {
96 content: String,
97 },
98 Tool {
99 tool_name: String,
100 content: String,
101 },
102}
103
104#[derive(Serialize, Deserialize, Debug)]
105#[serde(rename_all = "lowercase")]
106pub enum OllamaToolCall {
107 Function(OllamaFunctionCall),
108}
109
110#[derive(Serialize, Deserialize, Debug)]
111pub struct OllamaFunctionCall {
112 pub name: String,
113 pub arguments: Value,
114}
115
116#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
117pub struct OllamaFunctionTool {
118 pub name: String,
119 pub description: Option<String>,
120 pub parameters: Option<Value>,
121}
122
123#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
124#[serde(tag = "type", rename_all = "lowercase")]
125pub enum OllamaTool {
126 Function { function: OllamaFunctionTool },
127}
128
129#[derive(Serialize, Debug)]
130pub struct ChatRequest {
131 pub model: String,
132 pub messages: Vec<ChatMessage>,
133 pub stream: bool,
134 pub keep_alive: KeepAlive,
135 pub options: Option<ChatOptions>,
136 pub tools: Vec<OllamaTool>,
137 pub think: Option<bool>,
138}
139
140impl ChatRequest {
141 pub fn with_tools(mut self, tools: Vec<OllamaTool>) -> Self {
142 self.stream = false;
143 self.tools = tools;
144 self
145 }
146}
147
148// https://github.com/ollama/ollama/blob/main/docs/modelfile.md#valid-parameters-and-values
149#[derive(Serialize, Default, Debug)]
150pub struct ChatOptions {
151 pub num_ctx: Option<u64>,
152 pub num_predict: Option<isize>,
153 pub stop: Option<Vec<String>>,
154 pub temperature: Option<f32>,
155 pub top_p: Option<f32>,
156}
157
158#[derive(Deserialize, Debug)]
159pub struct ChatResponseDelta {
160 #[allow(unused)]
161 pub model: String,
162 #[allow(unused)]
163 pub created_at: String,
164 pub message: ChatMessage,
165 #[allow(unused)]
166 pub done_reason: Option<String>,
167 #[allow(unused)]
168 pub done: bool,
169 pub prompt_eval_count: Option<u64>,
170 pub eval_count: Option<u64>,
171}
172
173#[derive(Serialize, Deserialize)]
174pub struct LocalModelsResponse {
175 pub models: Vec<LocalModelListing>,
176}
177
178#[derive(Serialize, Deserialize)]
179pub struct LocalModelListing {
180 pub name: String,
181 pub modified_at: String,
182 pub size: u64,
183 pub digest: String,
184 pub details: ModelDetails,
185}
186
187#[derive(Serialize, Deserialize)]
188pub struct LocalModel {
189 pub modelfile: String,
190 pub parameters: String,
191 pub template: String,
192 pub details: ModelDetails,
193}
194
195#[derive(Serialize, Deserialize)]
196pub struct ModelDetails {
197 pub format: String,
198 pub family: String,
199 pub families: Option<Vec<String>>,
200 pub parameter_size: String,
201 pub quantization_level: String,
202}
203
204#[derive(Deserialize, Debug)]
205pub struct ModelShow {
206 #[serde(default)]
207 pub capabilities: Vec<String>,
208}
209
210impl ModelShow {
211 pub fn supports_tools(&self) -> bool {
212 // .contains expects &String, which would require an additional allocation
213 self.capabilities.iter().any(|v| v == "tools")
214 }
215
216 pub fn supports_vision(&self) -> bool {
217 self.capabilities.iter().any(|v| v == "vision")
218 }
219
220 pub fn supports_thinking(&self) -> bool {
221 self.capabilities.iter().any(|v| v == "thinking")
222 }
223}
224
225pub async fn complete(
226 client: &dyn HttpClient,
227 api_url: &str,
228 request: ChatRequest,
229) -> Result<ChatResponseDelta> {
230 let uri = format!("{api_url}/api/chat");
231 let request_builder = HttpRequest::builder()
232 .method(Method::POST)
233 .uri(uri)
234 .header("Content-Type", "application/json");
235
236 let serialized_request = serde_json::to_string(&request)?;
237 let request = request_builder.body(AsyncBody::from(serialized_request))?;
238
239 let mut response = client.send(request).await?;
240
241 let mut body = Vec::new();
242 response.body_mut().read_to_end(&mut body).await?;
243
244 if response.status().is_success() {
245 let response_message: ChatResponseDelta = serde_json::from_slice(&body)?;
246 Ok(response_message)
247 } else {
248 let body_str = std::str::from_utf8(&body)?;
249 anyhow::bail!(
250 "Failed to connect to API: {} {}",
251 response.status(),
252 body_str
253 );
254 }
255}
256
257pub async fn stream_chat_completion(
258 client: &dyn HttpClient,
259 api_url: &str,
260 request: ChatRequest,
261) -> Result<BoxStream<'static, Result<ChatResponseDelta>>> {
262 let uri = format!("{api_url}/api/chat");
263 let request_builder = http::Request::builder()
264 .method(Method::POST)
265 .uri(uri)
266 .header("Content-Type", "application/json");
267
268 let request = request_builder.body(AsyncBody::from(serde_json::to_string(&request)?))?;
269 let mut response = client.send(request).await?;
270 if response.status().is_success() {
271 let reader = BufReader::new(response.into_body());
272
273 Ok(reader
274 .lines()
275 .map(|line| match line {
276 Ok(line) => serde_json::from_str(&line).context("Unable to parse chat response"),
277 Err(e) => Err(e.into()),
278 })
279 .boxed())
280 } else {
281 let mut body = String::new();
282 response.body_mut().read_to_string(&mut body).await?;
283 anyhow::bail!(
284 "Failed to connect to Ollama API: {} {}",
285 response.status(),
286 body,
287 );
288 }
289}
290
291pub async fn get_models(
292 client: &dyn HttpClient,
293 api_url: &str,
294 _: Option<Duration>,
295) -> Result<Vec<LocalModelListing>> {
296 let uri = format!("{api_url}/api/tags");
297 let request_builder = HttpRequest::builder()
298 .method(Method::GET)
299 .uri(uri)
300 .header("Accept", "application/json");
301
302 let request = request_builder.body(AsyncBody::default())?;
303
304 let mut response = client.send(request).await?;
305
306 let mut body = String::new();
307 response.body_mut().read_to_string(&mut body).await?;
308
309 anyhow::ensure!(
310 response.status().is_success(),
311 "Failed to connect to Ollama API: {} {}",
312 response.status(),
313 body,
314 );
315 let response: LocalModelsResponse =
316 serde_json::from_str(&body).context("Unable to parse Ollama tag listing")?;
317 Ok(response.models)
318}
319
320/// Fetch details of a model, used to determine model capabilities
321pub async fn show_model(client: &dyn HttpClient, api_url: &str, model: &str) -> Result<ModelShow> {
322 let uri = format!("{api_url}/api/show");
323 let request = HttpRequest::builder()
324 .method(Method::POST)
325 .uri(uri)
326 .header("Content-Type", "application/json")
327 .body(AsyncBody::from(
328 serde_json::json!({ "model": model }).to_string(),
329 ))?;
330
331 let mut response = client.send(request).await?;
332 let mut body = String::new();
333 response.body_mut().read_to_string(&mut body).await?;
334
335 anyhow::ensure!(
336 response.status().is_success(),
337 "Failed to connect to Ollama API: {} {}",
338 response.status(),
339 body,
340 );
341 let details: ModelShow = serde_json::from_str(body.as_str())?;
342 Ok(details)
343}
344
345#[cfg(test)]
346mod tests {
347 use super::*;
348
349 #[test]
350 fn parse_completion() {
351 let response = serde_json::json!({
352 "model": "llama3.2",
353 "created_at": "2023-12-12T14:13:43.416799Z",
354 "message": {
355 "role": "assistant",
356 "content": "Hello! How are you today?"
357 },
358 "done": true,
359 "total_duration": 5191566416u64,
360 "load_duration": 2154458,
361 "prompt_eval_count": 26,
362 "prompt_eval_duration": 383809000,
363 "eval_count": 298,
364 "eval_duration": 4799921000u64
365 });
366 let _: ChatResponseDelta = serde_json::from_value(response).unwrap();
367 }
368
369 #[test]
370 fn parse_streaming_completion() {
371 let partial = serde_json::json!({
372 "model": "llama3.2",
373 "created_at": "2023-08-04T08:52:19.385406455-07:00",
374 "message": {
375 "role": "assistant",
376 "content": "The",
377 "images": null
378 },
379 "done": false
380 });
381
382 let _: ChatResponseDelta = serde_json::from_value(partial).unwrap();
383
384 let last = serde_json::json!({
385 "model": "llama3.2",
386 "created_at": "2023-08-04T19:22:45.499127Z",
387 "message": {
388 "role": "assistant",
389 "content": ""
390 },
391 "done": true,
392 "total_duration": 4883583458u64,
393 "load_duration": 1334875,
394 "prompt_eval_count": 26,
395 "prompt_eval_duration": 342546000,
396 "eval_count": 282,
397 "eval_duration": 4535599000u64
398 });
399
400 let _: ChatResponseDelta = serde_json::from_value(last).unwrap();
401 }
402
403 #[test]
404 fn parse_tool_call() {
405 let response = serde_json::json!({
406 "model": "llama3.2:3b",
407 "created_at": "2025-04-28T20:02:02.140489Z",
408 "message": {
409 "role": "assistant",
410 "content": "",
411 "tool_calls": [
412 {
413 "function": {
414 "name": "weather",
415 "arguments": {
416 "city": "london",
417 }
418 }
419 }
420 ]
421 },
422 "done_reason": "stop",
423 "done": true,
424 "total_duration": 2758629166u64,
425 "load_duration": 1770059875,
426 "prompt_eval_count": 147,
427 "prompt_eval_duration": 684637583,
428 "eval_count": 16,
429 "eval_duration": 302561917,
430 });
431
432 let result: ChatResponseDelta = serde_json::from_value(response).unwrap();
433 match result.message {
434 ChatMessage::Assistant {
435 content,
436 tool_calls,
437 images: _,
438 thinking,
439 } => {
440 assert!(content.is_empty());
441 assert!(tool_calls.is_some_and(|v| !v.is_empty()));
442 assert!(thinking.is_none());
443 }
444 _ => panic!("Deserialized wrong role"),
445 }
446 }
447
448 #[test]
449 fn parse_show_model() {
450 let response = serde_json::json!({
451 "license": "LLAMA 3.2 COMMUNITY LICENSE AGREEMENT...",
452 "details": {
453 "parent_model": "",
454 "format": "gguf",
455 "family": "llama",
456 "families": ["llama"],
457 "parameter_size": "3.2B",
458 "quantization_level": "Q4_K_M"
459 },
460 "model_info": {
461 "general.architecture": "llama",
462 "general.basename": "Llama-3.2",
463 "general.file_type": 15,
464 "general.finetune": "Instruct",
465 "general.languages": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
466 "general.parameter_count": 3212749888u64,
467 "general.quantization_version": 2,
468 "general.size_label": "3B",
469 "general.tags": ["facebook", "meta", "pytorch", "llama", "llama-3", "text-generation"],
470 "general.type": "model",
471 "llama.attention.head_count": 24,
472 "llama.attention.head_count_kv": 8,
473 "llama.attention.key_length": 128,
474 "llama.attention.layer_norm_rms_epsilon": 0.00001,
475 "llama.attention.value_length": 128,
476 "llama.block_count": 28,
477 "llama.context_length": 131072,
478 "llama.embedding_length": 3072,
479 "llama.feed_forward_length": 8192,
480 "llama.rope.dimension_count": 128,
481 "llama.rope.freq_base": 500000,
482 "llama.vocab_size": 128256,
483 "tokenizer.ggml.bos_token_id": 128000,
484 "tokenizer.ggml.eos_token_id": 128009,
485 "tokenizer.ggml.merges": null,
486 "tokenizer.ggml.model": "gpt2",
487 "tokenizer.ggml.pre": "llama-bpe",
488 "tokenizer.ggml.token_type": null,
489 "tokenizer.ggml.tokens": null
490 },
491 "tensors": [
492 { "name": "rope_freqs.weight", "type": "F32", "shape": [64] },
493 { "name": "token_embd.weight", "type": "Q4_K_S", "shape": [3072, 128256] }
494 ],
495 "capabilities": ["completion", "tools"],
496 "modified_at": "2025-04-29T21:24:41.445877632+03:00"
497 });
498
499 let result: ModelShow = serde_json::from_value(response).unwrap();
500 assert!(result.supports_tools());
501 assert!(result.capabilities.contains(&"tools".to_string()));
502 assert!(result.capabilities.contains(&"completion".to_string()));
503 }
504
505 #[test]
506 fn serialize_chat_request_with_images() {
507 let base64_image = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==";
508
509 let request = ChatRequest {
510 model: "llava".to_string(),
511 messages: vec![ChatMessage::User {
512 content: "What do you see in this image?".to_string(),
513 images: Some(vec![base64_image.to_string()]),
514 }],
515 stream: false,
516 keep_alive: KeepAlive::default(),
517 options: None,
518 think: None,
519 tools: vec![],
520 };
521
522 let serialized = serde_json::to_string(&request).unwrap();
523 assert!(serialized.contains("images"));
524 assert!(serialized.contains(base64_image));
525 }
526
527 #[test]
528 fn serialize_chat_request_without_images() {
529 let request = ChatRequest {
530 model: "llama3.2".to_string(),
531 messages: vec![ChatMessage::User {
532 content: "Hello, world!".to_string(),
533 images: None,
534 }],
535 stream: false,
536 keep_alive: KeepAlive::default(),
537 options: None,
538 think: None,
539 tools: vec![],
540 };
541
542 let serialized = serde_json::to_string(&request).unwrap();
543 assert!(!serialized.contains("images"));
544 }
545
546 #[test]
547 fn test_json_format_with_images() {
548 let base64_image = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==";
549
550 let request = ChatRequest {
551 model: "llava".to_string(),
552 messages: vec![ChatMessage::User {
553 content: "What do you see?".to_string(),
554 images: Some(vec![base64_image.to_string()]),
555 }],
556 stream: false,
557 keep_alive: KeepAlive::default(),
558 options: None,
559 think: None,
560 tools: vec![],
561 };
562
563 let serialized = serde_json::to_string(&request).unwrap();
564
565 let parsed: serde_json::Value = serde_json::from_str(&serialized).unwrap();
566 let message_images = parsed["messages"][0]["images"].as_array().unwrap();
567 assert_eq!(message_images.len(), 1);
568 assert_eq!(message_images[0].as_str().unwrap(), base64_image);
569 }
570}