switchyard_protocol/
llm.rs

1// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2// SPDX-License-Identifier: Apache-2.0
3
4//! Provider-neutral conversation types shared by routing, clients, and translation.
5
6use std::collections::BTreeMap;
7
8use serde::{Deserialize, Serialize};
9use serde_json::{Map, Value};
10
11use crate::format::FormatId;
12
13/// Actor role normalized across provider APIs.
14#[derive(Clone, Copy, Debug, Eq, PartialEq, Serialize, Deserialize)]
15#[serde(rename_all = "snake_case")]
16pub enum Role {
17    /// System-level instructions.
18    System,
19    /// Developer-level instructions supported by some provider APIs.
20    Developer,
21    /// End-user input.
22    User,
23    /// Model-generated output.
24    Assistant,
25    /// Tool execution output.
26    Tool,
27}
28
29/// Instruction content separated from normal conversation messages.
30#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
31pub struct InstructionBlock {
32    /// Instruction authority, normally [`Role::System`] or [`Role::Developer`].
33    pub role: Role,
34    /// Ordered instruction content.
35    pub content: Vec<ContentBlock>,
36}
37
38/// One normalized conversation message.
39#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
40pub struct Message {
41    /// Actor that produced the message.
42    pub role: Role,
43    /// Ordered message content.
44    pub content: Vec<ContentBlock>,
45}
46
47impl Message {
48    /// Creates a text-only message for the given role.
49    pub fn text(role: Role, text: impl Into<String>) -> Self {
50        Self {
51            role,
52            content: vec![ContentBlock::Text { text: text.into() }],
53        }
54    }
55
56    /// Concatenates text-like content blocks when the message has any.
57    pub fn text_content(&self, separator: &str) -> Option<String> {
58        let parts = self
59            .content
60            .iter()
61            .filter_map(|block| match block {
62                ContentBlock::Text { text } => Some(text.as_str()),
63                ContentBlock::Refusal { text } => Some(text.as_str()),
64                _ => None,
65            })
66            .collect::<Vec<_>>();
67        if parts.is_empty() {
68            None
69        } else {
70            Some(parts.join(separator))
71        }
72    }
73}
74
75/// Normalized content block variants carried by messages and tool results.
76#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
77#[serde(tag = "type", rename_all = "snake_case")]
78pub enum ContentBlock {
79    /// Plain text content.
80    Text {
81        /// Text value.
82        text: String,
83    },
84    /// Model reasoning or thinking content.
85    Reasoning {
86        /// Reasoning text.
87        text: String,
88        /// Provider signature used to validate or continue the reasoning block.
89        signature: Option<String>,
90        /// Structured reasoning details, such as an encrypted `{ "type":
91        /// "reasoning.encrypted", "data": "..." }` object, replayed without modification.
92        #[serde(default, skip_serializing_if = "Vec::is_empty")]
93        details: Vec<Value>,
94    },
95    /// Image content.
96    Image {
97        /// Image location or inline payload.
98        source: ImageSource,
99    },
100    /// Audio content.
101    Audio {
102        /// Audio location or inline payload.
103        source: MediaSource,
104    },
105    /// Video content.
106    Video {
107        /// Video location or inline payload.
108        source: MediaSource,
109    },
110    /// File content.
111    File {
112        /// File identifier or inline payload.
113        source: FileSource,
114    },
115    /// Tool invocation requested by the assistant.
116    ToolCall(ToolCall),
117    /// Result of an earlier tool invocation.
118    ToolResult(ToolResult),
119    /// Provider refusal content.
120    Refusal {
121        /// Human-readable refusal text.
122        text: String,
123    },
124    /// Provider block that has no normalized representation.
125    Unknown {
126        /// Wire format that supplied the block.
127        provider: FormatId,
128        /// Exact provider block.
129        raw: Value,
130    },
131}
132
133/// Image payload forms supported by the conversation model.
134#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
135#[serde(tag = "type", content = "data", rename_all = "snake_case")]
136pub enum ImageSource {
137    /// Image fetched from a URL.
138    Url {
139        /// Image URL.
140        url: String,
141        /// Optional provider-specific image detail level.
142        detail: Option<String>,
143    },
144    /// Base64-encoded image bytes.
145    Base64 {
146        /// MIME type, when supplied.
147        media_type: Option<String>,
148        /// Base64-encoded bytes.
149        data: String,
150    },
151    /// Provider image source with no normalized representation.
152    Raw(Value),
153}
154
155/// File payload forms supported by the conversation model.
156#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
157#[serde(tag = "type", content = "data", rename_all = "snake_case")]
158pub enum FileSource {
159    /// Provider-managed file identifier.
160    FileId(String),
161    /// Inline file payload.
162    FileData {
163        /// Encoded or textual file data as supplied by the provider format.
164        data: String,
165        /// Original filename, when supplied.
166        filename: Option<String>,
167    },
168    /// Provider file source with no normalized representation.
169    Raw(Value),
170}
171
172/// Audio and video payload forms supported by the conversation model.
173#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
174#[serde(tag = "type", content = "data", rename_all = "snake_case")]
175pub enum MediaSource {
176    /// Media fetched from a URL.
177    Url {
178        /// Media URL.
179        url: String,
180        /// MIME type, when supplied.
181        media_type: Option<String>,
182    },
183    /// Base64-encoded media bytes.
184    Base64 {
185        /// MIME type, when supplied.
186        media_type: Option<String>,
187        /// Base64-encoded bytes.
188        data: String,
189    },
190    /// Provider media source with no normalized representation.
191    Raw(Value),
192}
193
194/// Normalized assistant tool call.
195#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
196pub struct ToolCall {
197    /// Provider tool-call identifier used to pair the result.
198    pub id: String,
199    /// Tool name.
200    pub name: String,
201    /// Parsed tool arguments.
202    pub arguments: Value,
203}
204
205/// Normalized tool result message content.
206#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
207pub struct ToolResult {
208    /// Identifier of the [`ToolCall`] this result answers.
209    pub tool_call_id: String,
210    /// Ordered tool output content.
211    pub content: Vec<ContentBlock>,
212    /// Whether tool execution failed, when the provider reports it.
213    pub is_error: Option<bool>,
214}
215
216/// Normalized tool definition.
217#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
218pub struct ToolDefinition {
219    /// Tool name exposed to the model.
220    pub name: String,
221    /// Human-readable tool description.
222    pub description: Option<String>,
223    /// JSON Schema describing accepted arguments.
224    pub parameters: Value,
225    /// Whether the provider should enforce the schema strictly.
226    pub strict: Option<bool>,
227}
228
229/// Normalized tool choice policy.
230#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
231#[serde(tag = "type", content = "data", rename_all = "snake_case")]
232pub enum ToolChoice {
233    /// Let the model decide whether to call a tool.
234    Auto,
235    /// Require at least one tool call.
236    Required,
237    /// Disallow tool calls.
238    None,
239    /// Require a specific tool.
240    Tool {
241        /// Required tool name.
242        name: String,
243    },
244    /// Provider tool-choice value with no normalized representation.
245    Raw(Value),
246}
247
248/// Provider sampling parameters with common cross-provider names.
249#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
250pub struct SamplingParams {
251    /// Sampling temperature.
252    pub temperature: Option<f64>,
253    /// Nucleus-sampling probability mass.
254    pub top_p: Option<f64>,
255    /// Maximum number of candidate tokens considered at each step.
256    pub top_k: Option<i64>,
257}
258
259/// Output budget and structured-output options.
260#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
261pub struct OutputParams {
262    /// Maximum tokens the model may generate.
263    pub max_output_tokens: Option<u64>,
264    /// Provider-neutral or provider-specific structured-output configuration.
265    pub response_format: Option<Value>,
266}
267
268/// Provider reasoning controls preserved by translation.
269#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
270pub struct ReasoningParams {
271    /// Requested reasoning effort or level.
272    pub effort: Option<String>,
273    /// Provider reasoning controls without a normalized field.
274    pub raw: Option<Value>,
275}
276
277/// Provider-specific fields that do not have first-class conversation fields.
278#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
279#[serde(default)]
280pub struct ProviderExtensions {
281    /// Provider fields keyed by their original wire names.
282    ///
283    /// Codecs use these fields when translating to another format. Exact
284    /// same-format replay is controlled separately by [`PreservationMetadata`].
285    pub fields: Map<String, Value>,
286}
287
288/// Exact source payloads retained for lossless same-format round trips.
289///
290/// Translation's default preservation policy prefers a stored same-format body
291/// over reconstructing one from normalized fields. A caller that mutates the IR
292/// must clear the corresponding entry or use a policy with preservation disabled
293/// when those mutations must be encoded.
294#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
295#[serde(default)]
296pub struct PreservationMetadata {
297    /// Original request bodies keyed by source format.
298    pub requests: BTreeMap<FormatId, Value>,
299    /// Original response bodies keyed by source format.
300    pub responses: BTreeMap<FormatId, Value>,
301}
302
303/// Normalized request representation shared by Switchyard components.
304#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
305#[serde(default)]
306pub struct LlmRequest {
307    /// Model currently addressed by the request.
308    ///
309    /// This initially contains the name supplied by the inbound client. A routing host may
310    /// replace it with the selected target before serving the request.
311    pub model: Option<String>,
312    /// System and developer instructions separated from conversation turns.
313    pub instructions: Vec<InstructionBlock>,
314    /// Ordered conversation messages.
315    pub messages: Vec<Message>,
316    /// Tools available to the model.
317    pub tools: Vec<ToolDefinition>,
318    /// Policy controlling model tool selection.
319    pub tool_choice: Option<ToolChoice>,
320    /// Common sampling controls.
321    pub sampling: SamplingParams,
322    /// Output budget and shape controls.
323    pub output: OutputParams,
324    /// Reasoning controls.
325    pub reasoning: ReasoningParams,
326    /// Whether the caller requested a streamed response.
327    pub stream: bool,
328    /// Provider fields without first-class normalized equivalents.
329    pub extensions: ProviderExtensions,
330    /// Exact provider bodies used by codecs for lossless same-format round trips.
331    /// This is separate from a host's optional
332    /// [`Request::raw_request`](crate::Request::raw_request).
333    pub preservation: PreservationMetadata,
334}
335
336/// Normalized token usage counts.
337#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
338pub struct Usage {
339    /// Non-cached input tokens. Provider codecs normalize aggregate OpenAI
340    /// input counts by subtracting the cache detail fields.
341    pub input_tokens: Option<u64>,
342    /// Cache-read and cache-creation token detail, when reported.
343    #[serde(flatten)]
344    pub cache: Option<Box<InputCacheUsage>>,
345    /// Generated output tokens, excluding reasoning detail when the provider reports it separately.
346    pub output_tokens: Option<u64>,
347    /// Provider-reported or codec-computed total token count.
348    ///
349    /// Codecs normalize OpenAI aggregate input counts before computing totals, so
350    /// this may equal non-cached input plus cache detail plus output.
351    pub total_tokens: Option<u64>,
352    /// Reasoning output tokens, when reported separately.
353    pub reasoning_tokens: Option<u64>,
354}
355
356/// Optional cache-token detail kept out of the common, cache-free usage allocation.
357#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
358pub struct InputCacheUsage {
359    /// Input tokens read from a provider cache.
360    pub cached_input_tokens: Option<u64>,
361    /// Input tokens written into a provider cache.
362    pub cache_creation_input_tokens: Option<u64>,
363}
364
365impl Usage {
366    /// Builds cache detail when at least one count is present.
367    pub fn cache_details(
368        cached_input_tokens: Option<u64>,
369        cache_creation_input_tokens: Option<u64>,
370    ) -> Option<Box<InputCacheUsage>> {
371        if cached_input_tokens.is_none() && cache_creation_input_tokens.is_none() {
372            return None;
373        }
374        Some(Box::new(InputCacheUsage {
375            cached_input_tokens,
376            cache_creation_input_tokens,
377        }))
378    }
379
380    /// Returns the cache-read input-token count.
381    pub fn cached_input_tokens(&self) -> Option<u64> {
382        self.cache
383            .as_ref()
384            .and_then(|cache| cache.cached_input_tokens)
385    }
386
387    /// Returns the cache-creation input-token count.
388    pub fn cache_creation_input_tokens(&self) -> Option<u64> {
389        self.cache
390            .as_ref()
391            .and_then(|cache| cache.cache_creation_input_tokens)
392    }
393
394    /// Sets the cache-read count, allocating cache detail when needed.
395    pub fn set_cached_input_tokens(&mut self, value: u64) {
396        self.cache
397            .get_or_insert_with(|| Box::new(InputCacheUsage::default()))
398            .cached_input_tokens = Some(value);
399    }
400
401    /// Sets the cache-creation count, allocating cache detail when needed.
402    pub fn set_cache_creation_input_tokens(&mut self, value: u64) {
403        self.cache
404            .get_or_insert_with(|| Box::new(InputCacheUsage::default()))
405            .cache_creation_input_tokens = Some(value);
406    }
407}
408
409/// Normalized reason a model stopped producing output.
410#[derive(Clone, Copy, Debug, Eq, PartialEq, Serialize, Deserialize)]
411#[serde(rename_all = "snake_case")]
412pub enum StopReason {
413    /// Model completed its turn normally.
414    EndTurn,
415    /// Model reached the configured output limit.
416    MaxTokens,
417    /// Model stopped to request a tool call.
418    ToolUse,
419    /// Provider safety or content filtering stopped generation.
420    ContentFilter,
421    /// Generation terminated because of an error.
422    Error,
423    /// Provider stop reason with no normalized equivalent.
424    Unknown,
425}
426
427/// One assistant output item in a normalized response.
428#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
429pub struct ResponseOutput {
430    /// Actor that produced the output, normally [`Role::Assistant`].
431    pub role: Role,
432    /// Ordered output content.
433    pub content: Vec<ContentBlock>,
434    /// Why this output stopped, when known.
435    pub stop_reason: Option<StopReason>,
436}
437
438/// Normalized, fully-buffered response — the aggregate of a completed generation.
439/// This is the terminal form of a streamed [`LlmResponse`](crate::LlmResponse).
440#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
441#[serde(default)]
442pub struct AggLlmResponse {
443    /// Provider response identifier.
444    pub id: Option<String>,
445    /// Model reported by the provider.
446    pub model: Option<String>,
447    /// Ordered response output items.
448    pub outputs: Vec<ResponseOutput>,
449    /// Normalized token usage.
450    pub usage: Usage,
451    /// Provider response fields without normalized equivalents.
452    pub extensions: ProviderExtensions,
453    /// Exact provider bodies retained for lossless round trips.
454    pub preservation: PreservationMetadata,
455}
456
457impl AggLlmResponse {
458    /// Returns the first output item when a response has any output.
459    pub fn first_output(&self) -> Option<&ResponseOutput> {
460        self.outputs.first()
461    }
462}
463
464#[cfg(test)]
465mod tests {
466    use serde_json::json;
467
468    use super::*;
469
470    #[test]
471    fn serde_uses_python_friendly_dictionary_shapes() -> Result<(), serde_json::Error> {
472        let request: LlmRequest = serde_json::from_value(json!({
473            "model": "auto",
474            "messages": [{
475                "role": "user",
476                "content": [{"type": "text", "text": "hello"}]
477            }]
478        }))?;
479        assert_eq!(request.messages[0], Message::text(Role::User, "hello"));
480
481        let tool_call = ContentBlock::ToolCall(ToolCall {
482            id: "call-1".to_string(),
483            name: "lookup".to_string(),
484            arguments: json!({"query": "rust"}),
485        });
486        assert_eq!(
487            serde_json::to_value(tool_call)?,
488            json!({
489                "type": "tool_call",
490                "id": "call-1",
491                "name": "lookup",
492                "arguments": {"query": "rust"}
493            })
494        );
495        Ok(())
496    }
497}
switchyard_protocol/llm.rs

switchyard_protocol/
llm.rs
