1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
//! Google Vertex AI provider.
//!
//! Similar to Google Generative AI but uses OAuth2 authentication
//! and a different base URL pattern with project/location.
//!
//! The API key in StreamConfig is expected to be an OAuth2 access token.
//! Callers are responsible for obtaining the token (e.g., via service account JWT).
use super::model::ModelConfig;
use super::traits::*;
use crate::types::*;
use async_trait::async_trait;
use tokio::sync::mpsc;
pub struct GoogleVertexProvider;
impl GoogleVertexProvider {
/// Build the Vertex AI URL from model config.
/// Expects base_url in format: `https://{region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/publishers/google/models`
fn vertex_url(model_config: &ModelConfig, model: &str) -> String {
format!(
"{}/{}:streamGenerateContent?alt=sse",
model_config.base_url, model
)
}
}
#[async_trait]
impl StreamProvider for GoogleVertexProvider {
async fn stream(
&self,
config: StreamConfig,
tx: mpsc::UnboundedSender<StreamEvent>,
cancel: tokio_util::sync::CancellationToken,
) -> Result<Message, ProviderError> {
let model_config = config
.model_config
.as_ref()
.ok_or_else(|| ProviderError::Other("ModelConfig required".into()))?;
// Override the base_url to use Vertex format.
// The GoogleProvider's stream will use model_config.base_url, but we need
// a different URL pattern. We delegate to GoogleProvider with a modified config.
let vertex_url = Self::vertex_url(model_config, &config.model);
// Create a modified model config with the Vertex URL pattern
let mut vertex_model = model_config.clone();
// For Vertex, auth is via Bearer token (OAuth2), not API key in query param.
// We need to add the Authorization header.
vertex_model.headers.insert(
"authorization".to_string(),
format!("Bearer {}", config.api_key),
);
// Build request body same as Google (same content format)
let body = build_vertex_request_body(&config);
let client = reqwest::Client::new();
let mut request = client
.post(&vertex_url)
.header("content-type", "application/json");
for (k, v) in &vertex_model.headers {
request = request.header(k, v);
}
let response = request
.json(&body)
.send()
.await
.map_err(|e| ProviderError::Network(e.to_string()))?;
if !response.status().is_success() {
let status = response.status();
let body = response.text().await.unwrap_or_default();
return Err(ProviderError::classify(
status.as_u16(),
&format!("Vertex AI error {}: {}", status, body),
));
}
// Delegate SSE parsing to the Google provider's streaming logic.
// Since the response format is identical, we reuse GoogleProvider.
// However, we already have the response, so we'll parse it inline.
// For simplicity, delegate fully to GoogleProvider with modified config.
let mut modified_config = config.clone();
modified_config.model_config = Some(vertex_model);
// Actually, let's just delegate to GoogleProvider. The key difference
// is auth (Bearer vs API key in URL). We handle that by using a modified
// model config. But GoogleProvider builds its own URL... so let's just
// use GoogleProvider with a trick: empty api_key and auth in headers.
// We can't easily reuse GoogleProvider because it constructs its own URL.
// Instead, parse the SSE response directly (same format as Google GenAI).
parse_google_sse_response(response, &config, &model_config.provider, tx, cancel).await
}
}
/// Parse a Google-format SSE response stream. Shared between Google and Vertex.
async fn parse_google_sse_response(
response: reqwest::Response,
config: &StreamConfig,
provider_name: &str,
tx: mpsc::UnboundedSender<StreamEvent>,
cancel: tokio_util::sync::CancellationToken,
) -> Result<Message, ProviderError> {
use futures::StreamExt;
use serde::Deserialize;
use tracing::{debug, warn};
let mut content: Vec<Content> = Vec::new();
let mut usage = Usage::default();
let mut stop_reason = StopReason::Stop;
let _ = tx.send(StreamEvent::Start);
let mut stream = response.bytes_stream();
let mut buffer = String::new();
loop {
tokio::select! {
_ = cancel.cancelled() => {
return Err(ProviderError::Cancelled);
}
chunk = stream.next() => {
match chunk {
None => break,
Some(Err(e)) => {
warn!("Vertex stream error: {}", e);
break;
}
Some(Ok(bytes)) => {
buffer.push_str(&String::from_utf8_lossy(&bytes));
while let Some(pos) = buffer.find("\n\n") {
let event_str = buffer[..pos].to_string();
buffer = buffer[pos + 2..].to_string();
let data = event_str
.lines()
.find(|l| l.starts_with("data: "))
.map(|l| &l[6..])
.unwrap_or("");
if data.is_empty() {
continue;
}
#[derive(Deserialize)]
struct Chunk {
#[serde(default)]
candidates: Option<Vec<Candidate>>,
#[serde(default, rename = "usageMetadata")]
usage_metadata: Option<UsageMeta>,
}
#[derive(Deserialize)]
struct Candidate {
#[serde(default)]
content: Option<CContent>,
#[serde(default, rename = "finishReason")]
finish_reason: Option<String>,
}
#[derive(Deserialize)]
struct CContent {
#[serde(default)]
parts: Vec<Part>,
}
#[derive(Deserialize)]
struct Part {
#[serde(default)]
text: Option<String>,
#[serde(default, rename = "functionCall")]
function_call: Option<FCall>,
}
#[derive(Deserialize)]
struct FCall {
name: String,
#[serde(default)]
args: Option<serde_json::Value>,
}
#[derive(Deserialize)]
struct UsageMeta {
#[serde(default, rename = "promptTokenCount")]
prompt_token_count: Option<u64>,
#[serde(default, rename = "candidatesTokenCount")]
candidates_token_count: Option<u64>,
#[serde(default, rename = "totalTokenCount")]
total_token_count: Option<u64>,
}
let parsed: Chunk = match serde_json::from_str(data) {
Ok(c) => c,
Err(e) => {
debug!("Failed to parse Vertex chunk: {}", e);
continue;
}
};
for candidate in parsed.candidates.unwrap_or_default() {
if let Some(c) = candidate.content {
for part in c.parts {
if let Some(text) = part.text {
let idx = content.iter().position(|c| matches!(c, Content::Text { .. }));
let idx = match idx {
Some(i) => i,
None => {
content.push(Content::Text { text: String::new() });
content.len() - 1
}
};
if let Some(Content::Text { text: t }) = content.get_mut(idx) {
t.push_str(&text);
}
let _ = tx.send(StreamEvent::TextDelta {
content_index: idx,
delta: text,
});
}
if let Some(fc) = part.function_call {
let id = format!("vertex-fc-{}", content.len());
let args = fc.args.unwrap_or(serde_json::Value::Object(Default::default()));
let idx = content.len();
content.push(Content::ToolCall { provider_metadata: None,
id: id.clone(),
name: fc.name.clone(),
arguments: args,
});
let _ = tx.send(StreamEvent::ToolCallStart {
content_index: idx,
id,
name: fc.name,
});
let _ = tx.send(StreamEvent::ToolCallEnd { content_index: idx });
stop_reason = StopReason::ToolUse;
}
}
}
if let Some(reason) = candidate.finish_reason {
stop_reason = match reason.as_str() {
"STOP" => StopReason::Stop,
"MAX_TOKENS" => StopReason::Length,
_ => StopReason::Stop,
};
}
}
if let Some(u) = parsed.usage_metadata {
usage.input = u.prompt_token_count.unwrap_or(0);
usage.output = u.candidates_token_count.unwrap_or(0);
usage.total_tokens = u.total_token_count.unwrap_or(0);
}
}
}
}
}
}
}
let message = Message::Assistant {
content,
stop_reason,
model: config.model.clone(),
provider: provider_name.to_string(),
usage,
timestamp: now_ms(),
error_message: None,
};
let _ = tx.send(StreamEvent::Done {
message: message.clone(),
});
Ok(message)
}
/// Build the request body for Vertex AI (same format as Google GenAI).
fn build_vertex_request_body(config: &StreamConfig) -> serde_json::Value {
// Same format as Google GenAI
let mut contents: Vec<serde_json::Value> = Vec::new();
for msg in &config.messages {
match msg {
Message::User { content, .. } => {
let parts: Vec<serde_json::Value> = content
.iter()
.filter(|c| !matches!(c, Content::Text { text } if text.is_empty()))
.filter_map(|c| match c {
Content::Text { text } => Some(serde_json::json!({"text": text})),
Content::Image { data, mime_type } => Some(serde_json::json!({
"inlineData": {"mimeType": mime_type, "data": data},
})),
_ => None,
})
.collect();
contents.push(serde_json::json!({"role": "user", "parts": parts}));
}
Message::Assistant { content, .. } => {
let parts: Vec<serde_json::Value> = content
.iter()
.filter(|c| !matches!(c, Content::Text { text } if text.is_empty()))
.filter_map(|c| match c {
Content::Text { text } => Some(serde_json::json!({"text": text})),
Content::ToolCall {
name, arguments, ..
} => Some(serde_json::json!({
"functionCall": {"name": name, "args": arguments},
})),
_ => None,
})
.collect();
contents.push(serde_json::json!({"role": "model", "parts": parts}));
}
Message::ToolResult {
tool_name, content, ..
} => {
let text = content
.iter()
.find_map(|c| match c {
Content::Text { text } => Some(text.clone()),
_ => None,
})
.unwrap_or_default();
let mut parts = vec![serde_json::json!({
"functionResponse": {"name": tool_name, "response": {"result": text}}
})];
for c in content {
if let Content::Image { data, mime_type } = c {
parts.push(serde_json::json!({
"inlineData": {"mimeType": mime_type, "data": data},
}));
}
}
contents.push(serde_json::json!({
"role": "user",
"parts": parts,
}));
}
}
}
let mut body = serde_json::json!({"contents": contents});
if !config.system_prompt.is_empty() {
body["systemInstruction"] = serde_json::json!({"parts": [{"text": config.system_prompt}]});
}
let mut gen_config = serde_json::json!({});
if let Some(max) = config.max_tokens {
gen_config["maxOutputTokens"] = serde_json::json!(max);
}
if let Some(temp) = config.temperature {
gen_config["temperature"] = serde_json::json!(temp);
}
if gen_config != serde_json::json!({}) {
body["generationConfig"] = gen_config;
}
if !config.tools.is_empty() {
let declarations: Vec<serde_json::Value> = config
.tools
.iter()
.map(|t| {
serde_json::json!({
"name": t.name,
"description": t.description,
"parameters": t.parameters,
})
})
.collect();
body["tools"] = serde_json::json!([{"functionDeclarations": declarations}]);
}
body
}