meridian/llms/
open_ai.rs

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
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
pub mod messages;

use super::{
    messages::AbstractMessage, LLMProvider, LLMToolUsage, MultiModelLLMProvider,
    StructuredLLMProvider, Tool, ToolChoice, Toolkit,
};
use anyhow::Result;
use log::{debug, info, warn};
use messages::OpenAIMessage;
use reqwest::blocking::Client;
use schemars::{
    schema::{ObjectValidation, RootSchema, Schema},
    schema_for, JsonSchema,
};
use serde::{Deserialize, Serialize};

pub struct OpenAIClient {
    api_key: String,
    client: Client,
    model: OpenAIModel,
}

#[derive(Debug, Serialize, Deserialize, Clone, Copy)]
pub enum OpenAIModel {
    #[serde(rename = "gpt-4o")]
    Gpt4o,
    #[serde(rename = "o1-preview")]
    O1Preview,
}

#[derive(Debug, Serialize, Deserialize)]
pub struct CompletionRequest {
    model: OpenAIModel,
    messages: Vec<OpenAIMessage>,
}

impl CompletionRequest {
    fn body(model: OpenAIModel, messages: Vec<OpenAIMessage>) -> Self {
        Self { model, messages }
    }
}

#[derive(Debug, Deserialize)]
pub struct CompletionChoice {
    finish_reason: String,
    index: u64,
    message: OpenAIMessage,
}

#[derive(Debug, Deserialize)]
pub struct CompletionResponse {
    id: String,
    object: String,
    created: u64, // unix timestamp
    choices: Vec<CompletionChoice>,
}

fn set_additional_properties_false(root_schema: &mut RootSchema) {
    // Set root schema
    if root_schema.schema.object.is_none() {
        root_schema.schema.object = Some(Box::new(ObjectValidation::default()));
    }
    root_schema
        .schema
        .object
        .as_mut()
        .unwrap()
        .additional_properties = Some(Box::new(Schema::Bool(false)));

    // Set for properties
    if let Some(props) = &mut root_schema.schema.object {
        for schema in props.properties.values_mut() {
            if let Schema::Object(obj) = schema {
                if obj.object.is_none() {
                    obj.object = Some(Box::new(ObjectValidation::default()));
                }
                obj.object.as_mut().unwrap().additional_properties =
                    Some(Box::new(Schema::Bool(false)));
            }
        }
    }

    // Set for definitions
    for schema in root_schema.definitions.values_mut() {
        if let Schema::Object(obj) = schema {
            if obj.object.is_none() {
                obj.object = Some(Box::new(ObjectValidation::default()));
            }
            obj.object.as_mut().unwrap().additional_properties =
                Some(Box::new(Schema::Bool(false)));
        }
    }
}

impl LLMProvider<OpenAIMessage> for OpenAIClient {
    fn get_completion(&self, messages: Vec<OpenAIMessage>) -> Result<Vec<OpenAIMessage>> {
        debug!(
            "Getting completion from OpenAI with {} messages",
            messages.len()
        );

        let mut headers = reqwest::header::HeaderMap::new();
        headers.insert(
            "Authorization",
            format!("Bearer {}", self.api_key)
                .parse()
                .expect("Invalid API key"),
        );
        headers.insert(
            "Content-Type",
            "application/json".parse().expect("Invalid content type"),
        );

        let request_body = CompletionRequest::body(OpenAIModel::Gpt4o, messages.clone());
        debug!("Sending request to OpenAI API");

        let result = self
            .client
            .post("https://api.openai.com/v1/chat/completions")
            .headers(headers)
            .json(&request_body)
            .send()?;

        if !result.status().is_success() {
            let status = result.status();
            let error_text = result.text()?;
            warn!("OpenAI API error: {} - {}", status, error_text);
            return Err(anyhow::anyhow!(
                "Failed to get completion: {:?} {:?}",
                status,
                error_text
            ));
        }

        let completion_response: CompletionResponse = result.json()?;

        let last_message = completion_response.choices.first().ok_or(anyhow::anyhow!(
            "No choices returned in the OpenAI response"
        ))?;
        debug!("Last message: {:?}", last_message.message);

        Ok(messages
            .into_iter()
            .chain(vec![last_message.message.clone()])
            .collect())
    }

    fn stream_completion(
        &self,
        messages: Vec<OpenAIMessage>,
    ) -> Result<Box<dyn Iterator<Item = OpenAIMessage>>> {
        todo!("Implement streaming for the OpenAI client")
    }
}

impl MultiModelLLMProvider<OpenAIModel> for OpenAIClient {
    fn with_model(&self, model: OpenAIModel) -> Self {
        Self {
            api_key: self.api_key.clone(),
            client: self.client.clone(),
            model,
        }
    }

    fn get_model(&self) -> OpenAIModel {
        self.model
    }
}

impl LLMToolUsage<OpenAIMessage> for OpenAIClient {
    fn do_work_with_tool(
        &self,
        messages: Vec<OpenAIMessage>,
        tool: &dyn Tool,
    ) -> Result<Vec<OpenAIMessage>> {
        debug!("Executing tool '{}' with OpenAI", tool.name());

        let mut headers = reqwest::header::HeaderMap::new();
        headers.insert(
            "Authorization",
            format!("Bearer {}", self.api_key).parse().unwrap(),
        );
        headers.insert(
            "Content-Type",
            "application/json".parse().expect("Invalid content type"),
        );

        let request_body = serde_json::json!({
            "model": self.model,
            "messages": messages,
            "tools": [{
                "type": "function",
                "function": {
                    "name": tool.name(),
                    "description": tool.description(),
                    "parameters": tool.schema()
                }
            }],
            "tool_choice": {
                "type": "function",
                "function": { "name": tool.name() }
            }
        });

        debug!("Sending tool execution request to OpenAI API");
        let result = self
            .client
            .post("https://api.openai.com/v1/chat/completions")
            .headers(headers)
            .json(&request_body)
            .send()?;

        if !result.status().is_success() {
            let status = result.status();
            let error_text = result.text()?;
            warn!(
                "OpenAI API error during tool execution: {} - {}",
                status, error_text
            );
            return Err(anyhow::anyhow!("Failed to use tool: {}", error_text));
        }

        let response: CompletionResponse = result.json()?;
        // Debugging: Print the raw response
        println!("Raw response from tool use ask: {:#?}", response);

        let message = response
            .choices
            .first()
            .ok_or_else(|| anyhow::anyhow!("No choices returned in the OpenAI response"))?;
        debug!("Last message: {:?}", message.message);

        match &message.message {
            OpenAIMessage::Assistant {
                tool_calls: Some(tool_calls),
                ..
            } => {
                let tool_call = tool_calls
                    .first()
                    .ok_or_else(|| anyhow::anyhow!("No tool calls in assistant message"))?;

                let args = serde_json::from_str(&tool_call.function.arguments)?;
                let result = tool.execute(args)?;

                Ok(vec![OpenAIMessage::Tool {
                    content: serde_json::to_string(&result)?,
                    tool_call_id: tool_call.id.clone(),
                }])
            }
            _ => Err(anyhow::anyhow!(
                "Expected assistant message with tool calls"
            )),
        }
    }

    fn get_chat_with_tools(
        &self,
        messages: Vec<OpenAIMessage>,
        tool_kit: &Toolkit,
        force_tool_use: &ToolChoice,
    ) -> Result<Vec<OpenAIMessage>> {
        let mut headers = reqwest::header::HeaderMap::new();
        headers.insert(
            "Authorization",
            format!("Bearer {}", self.api_key).parse().unwrap(),
        );
        headers.insert(
            "Content-Type",
            "application/json".parse().expect("Invalid content type"),
        );

        debug!("Messages: {:?}", messages);

        let tool_defs: Vec<serde_json::Value> = tool_kit
            .tools()
            .iter()
            .map(|tool| {
                serde_json::json!({
                    "type": "function",
                    "function": {
                        "name": tool.name(),
                        "description": tool.description(),
                        "parameters": tool.schema()
                    }
                })
            })
            .collect();

        debug!("Tool definitions: {:?}", tool_defs);

        let tool_choice = match force_tool_use {
            ToolChoice::Specific(name) => serde_json::json!({
                "type": "function",
                "function": {
                    "name": name
                }
            }),
            ToolChoice::Any => serde_json::json!("required"),
            ToolChoice::SelfSelect => serde_json::json!("auto"),
        };

        let request_body = serde_json::json!({
            "model": self.model,
            "messages": messages,
            "tools": tool_defs,
            "tool_choice": tool_choice
        });

        let result = self
            .client
            .post("https://api.openai.com/v1/chat/completions")
            .headers(headers)
            .json(&request_body)
            .send()?;

        if !result.status().is_success() {
            let status = result.status();
            let error_text = result.text()?;
            warn!(
                "OpenAI API error during chat with tools: {} - {}",
                status, error_text
            );
            return Err(anyhow::anyhow!("Failed to chat with tools: {}", error_text));
        }

        let response: CompletionResponse = result.json()?;

        let message = response
            .choices
            .first()
            .ok_or_else(|| anyhow::anyhow!("No choices returned in the OpenAI response"))?;
        debug!("Last message: {:?}", message.message);

        // Return all the messages
        Ok(messages
            .into_iter()
            .chain(vec![message.message.clone()])
            .collect())
    }

    fn get_work_result(
        &self,
        messages: Vec<OpenAIMessage>,
        tool_kit: &Toolkit,
        tool_choice: &ToolChoice,
    ) -> Result<Vec<OpenAIMessage>> {
        info!("Getting work result with tool choice: {:?}", tool_choice);

        match tool_choice {
            ToolChoice::Specific(name) => {
                debug!("Using specific tool: {}", name);
                self.do_work_with_tool(
                    messages,
                    tool_kit
                        .get(name)
                        .ok_or_else(|| anyhow::anyhow!("Tool not found: {}", name))?,
                )
            }
            ToolChoice::Any => {
                debug!("Getting chat with any tool allowed");
                let response = self.get_chat_with_tools(messages, tool_kit, tool_choice)?;
                debug!("Response from chat with tools: {:?}", response);

                if let Some(OpenAIMessage::Assistant {
                    tool_calls: Some(tool_calls),
                    ..
                }) = response.clone().last()
                {
                    let mut result_messages = response;

                    // Process all tool calls
                    for tool_call in tool_calls {
                        debug!("Processing tool call: {:?}", tool_call);
                        let tool = tool_kit.get(&tool_call.function.name).ok_or_else(|| {
                            anyhow::anyhow!("Tool not found: {}", tool_call.function.name)
                        })?;

                        let args = serde_json::from_str(&tool_call.function.arguments)?;
                        let result = tool.execute(args)?;

                        result_messages.push(OpenAIMessage::Tool {
                            content: serde_json::to_string(&result)?,
                            tool_call_id: tool_call.id.clone(),
                        });
                    }

                    debug!("Result messages: {:?}", result_messages);

                    let messages = self.get_work_result(result_messages, tool_kit, tool_choice)?;
                    Ok(messages)
                } else {
                    Err(anyhow::anyhow!("No tool calls in assistant message"))
                }
            }
            ToolChoice::SelfSelect => {
                debug!("Letting model select tool usage");
                let response = self.get_chat_with_tools(messages, tool_kit, tool_choice)?;
                debug!("Response from chat with tools: {:?}", response);

                if let Some(OpenAIMessage::Assistant {
                    tool_calls: Some(tool_calls),
                    ..
                }) = response.clone().last()
                {
                    let mut result_messages = response;

                    // Process all tool calls
                    for tool_call in tool_calls {
                        debug!("Processing tool call: {:?}", tool_call);
                        let tool = tool_kit.get(&tool_call.function.name).ok_or_else(|| {
                            anyhow::anyhow!("Tool not found: {}", tool_call.function.name)
                        })?;

                        let args = serde_json::from_str(&tool_call.function.arguments)?;
                        let result = tool.execute(args)?;

                        result_messages.push(OpenAIMessage::Tool {
                            content: serde_json::to_string(&result)?,
                            tool_call_id: tool_call.id.clone(),
                        });
                    }

                    debug!("Result messages: {:?}", result_messages);

                    let messages = self.get_work_result(result_messages, tool_kit, tool_choice)?;
                    Ok(messages)
                } else {
                    Ok(response) // For SelfSelect, we return the response even if no tools were used
                }
            }
        }
    }
}

impl StructuredLLMProvider<OpenAIMessage> for OpenAIClient {
    fn get_structured_response<
        DesiredSchema: Serialize + serde::de::DeserializeOwned + JsonSchema,
    >(
        &self,
        messages: Vec<OpenAIMessage>,
    ) -> Result<DesiredSchema> {
        let mut headers = reqwest::header::HeaderMap::new();
        headers.insert(
            "Authorization",
            format!("Bearer {}", self.api_key)
                .parse()
                .expect("Invalid API key"),
        );
        headers.insert(
            "Content-Type",
            "application/json".parse().expect("Invalid content type"),
        );

        let mut schema = schema_for!(DesiredSchema);
        set_additional_properties_false(&mut schema);

        println!("{}", serde_json::to_string(&schema).unwrap());

        let request_body = serde_json::json!({
            "model": OpenAIModel::Gpt4o,
            "messages": messages,
            "response_format": {
                "type": "json_schema",
                "json_schema": {
                    "name": "desired_schema",
                    "strict": true,
                    "schema": schema
                }
            }
        });

        let result = self
            .client
            .post("https://api.openai.com/v1/chat/completions")
            .headers(headers)
            .json(&request_body)
            .send()?;

        if !result.status().is_success() {
            return Err(anyhow::anyhow!(
                "Failed to get structured response: {:?} {:?}",
                result.status(),
                result.text()
            ));
        }

        let response: CompletionResponse = result.json()?;

        let content = response.choices[0]
            .message
            .get_content()
            .map_err(|_| anyhow::anyhow!("Failed to get message content"))?;

        Ok(serde_json::from_str(&content)?)
    }
}

impl Default for OpenAIClient {
    fn default() -> Self {
        Self::new()
    }
}

impl OpenAIClient {
    pub fn new() -> Self {
        Self {
            api_key: std::env::var("OPENAI_API_KEY").expect("OPENAI_API_KEY not set"),
            client: Client::new(),
            model: OpenAIModel::Gpt4o,
        }
    }
}