aither_core/llm/
mod.rs

1//! # Language Models and Conversation Management
2//!
3//! This module provides everything you need to work with language models in a provider-agnostic way.
4//! Build chat applications, generate structured output, and integrate tools without being tied to any specific AI service.
5//!
6//! ## Core Components
7//!
8//! - **[`LanguageModel`]** - The main trait for text generation and conversation
9//! - **[`LLMRequest`]** - Encapsulates messages, tools, and parameters for model calls
10//! - **[`Event`]** - Stream events from the model (text, reasoning, tool calls)
11//! - **[`Message`]** - Represents individual messages in a conversation
12//! - **[`Tool`]** - Function calling interface for extending model capabilities
13//!
14//! ## Design Philosophy
15//!
16//! The core crate provides a **low-level API** that emits events without executing tools.
17//! Tool execution is the responsibility of higher-level abstractions like `aither-agent`.
18//!
19//! This design allows:
20//! - Full control over tool execution flow
21//! - Hooks for intercepting and modifying tool calls
22//! - Proper context management between turns
23//! - Clean separation between LLM communication and agent logic
24//!
25//! ## Quick Start
26//!
27//! ### Basic Conversation
28//!
29//! ```rust,ignore
30//! use aither::llm::{LanguageModel, Event, oneshot};
31//! use futures_lite::StreamExt;
32//!
33//! async fn chat_with_model(model: impl LanguageModel) -> Result<String, Box<dyn std::error::Error>> {
34//!     let request = oneshot("You are a helpful assistant", "What's the capital of Japan?");
35//!     let mut stream = model.respond(request);
36//!     let mut full_text = String::new();
37//!
38//!     while let Some(event) = stream.next().await {
39//!         match event? {
40//!             Event::Text(chunk) => full_text.push_str(&chunk),
41//!             Event::Reasoning(thought) => eprintln!("[thinking] {}", thought),
42//!             Event::ToolCall(call) => {
43//!                 // Handle tool call (typically done by agent crate)
44//!                 println!("Tool requested: {}", call.name);
45//!             }
46//!             _ => {}
47//!         }
48//!     }
49//!
50//!     Ok(full_text)
51//! }
52//! ```
53//!
54//! ### With Tools (Agent-Controlled)
55//!
56//! ```rust,ignore
57//! use aither::llm::{LanguageModel, Event, LLMRequest, Message};
58//!
59//! // The core crate does NOT execute tools - it emits ToolCall events.
60//! // Tool execution should be handled by the agent crate.
61//! let request = LLMRequest::new([Message::user("What's the weather?")])
62//!     .with_tool_definitions(vec![weather_tool_definition()]);
63//!
64//! let mut stream = model.respond(request);
65//! while let Some(event) = stream.next().await {
66//!     match event? {
67//!         Event::ToolCall(call) => {
68//!             // Execute tool and continue conversation
69//!             let result = my_tool_executor.execute(&call).await;
70//!             // Add result to messages and send another request...
71//!         }
72//!         _ => {}
73//!     }
74//! }
75//! ```
76
77/// Assistant module for managing assistant-related functionality.
78pub mod assistant;
79/// Event types for streaming responses.
80pub mod event;
81/// Classification of provider failures.
82pub mod failure;
83/// Message types and conversation handling.
84pub mod message;
85/// Model profiles and capabilities.
86pub mod model;
87/// Provider module for managing language model providers and their configurations.
88pub mod provider;
89/// Provider-opaque reasoning state carried across turns.
90pub mod reasoning;
91/// Deep research workflows and agent capabilities.
92pub mod researcher;
93/// How a failed model request is retried.
94pub mod retry;
95/// Tool system for function calling.
96pub mod tool;
97
98use crate::llm::{model::Parameters, tool::Tools};
99use alloc::{
100    boxed::Box,
101    string::{String, ToString},
102    sync::Arc,
103    vec,
104    vec::Vec,
105};
106use core::{any::TypeId, future::Future};
107pub use event::{Event, ToolCall, Usage};
108pub use failure::{Classify, Failure, FailureKind};
109use futures_core::Stream;
110use futures_lite::{StreamExt, pin};
111pub use message::{Attachment, Message, Role};
112pub use provider::LanguageModelProvider;
113pub use reasoning::ReasoningState;
114pub use researcher::{
115    ResearchCitation, ResearchEvent, ResearchFinding, ResearchOptions, ResearchReport,
116    ResearchRequest, ResearchSource, ResearchStage, Researcher, ResearcherProfile,
117};
118pub use retry::RetryPolicy;
119use schemars::{JsonSchema, schema_for};
120use serde::de::DeserializeOwned;
121pub use tool::{
122    AgentIdentity, CallIdentity, CommandOutcome, CommandResult, CommandStream, Interruption,
123    InterruptionCause, IntoToolResult, StoredFile, SubagentIdentity, Tool, ToolCallId, ToolContext,
124    ToolResult, render_interruptions,
125};
126
127use crate::llm::model::Profile;
128
129/// Why a structured-output call failed.
130///
131/// [`LanguageModel::respond`] reports provider failures through the model's own
132/// [`LanguageModel::Error`]. Structured output adds a second, separate way to
133/// fail — the model answered, but not with the requested shape — so this keeps
134/// the two distinguishable instead of flattening both into one opaque error.
135/// Callers need that distinction: a rate limit is worth retrying, a schema
136/// mismatch is worth re-prompting, and an auth failure is worth neither.
137#[derive(Debug)]
138pub enum GenerateError<E> {
139    /// The provider itself failed: transport, rate limit, auth, and so on.
140    Provider(E),
141
142    /// The provider answered, but the response did not match the schema.
143    ///
144    /// Carries the text that could not be parsed so callers can log or retry.
145    Parse {
146        /// The underlying deserialization failure.
147        source: serde_json::Error,
148        /// The response text that failed to parse, truncated for logging.
149        response: String,
150    },
151}
152
153impl<E: core::fmt::Display> core::fmt::Display for GenerateError<E> {
154    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
155        match self {
156            Self::Provider(err) => write!(f, "language model request failed: {err}"),
157            Self::Parse { source, response } => {
158                write!(
159                    f,
160                    "structured output did not match the requested schema: {source}; response: {response}"
161                )
162            }
163        }
164    }
165}
166
167impl<E: core::error::Error + 'static> core::error::Error for GenerateError<E> {
168    fn source(&self) -> Option<&(dyn core::error::Error + 'static)> {
169        match self {
170            Self::Provider(err) => Some(err),
171            Self::Parse { source, .. } => Some(source),
172        }
173    }
174}
175
176/// Builder-style request passed into [`LanguageModel::respond`].
177///
178/// Wraps the full conversation, model parameters, and tool definitions a provider
179/// needs in order to execute a call.
180#[derive(Debug, Clone)]
181pub struct LLMRequest {
182    messages: Vec<Message>,
183    parameters: Parameters,
184    tool_definitions: Vec<tool::ToolDefinition>,
185}
186
187impl LLMRequest {
188    /// Creates a request from the provided messages using default parameters.
189    pub fn new(messages: impl Into<Vec<Message>>) -> Self {
190        Self {
191            messages: messages.into(),
192            parameters: Parameters::default(),
193            tool_definitions: Vec::new(),
194        }
195    }
196
197    /// Adds tool definitions to the request.
198    ///
199    /// These define what tools the model can request. The model will emit
200    /// [`Event::ToolCall`] events when it wants to use a tool.
201    #[must_use]
202    pub fn with_tool_definitions(mut self, definitions: Vec<tool::ToolDefinition>) -> Self {
203        self.tool_definitions = definitions;
204        self
205    }
206
207    /// Adds a single tool definition.
208    #[must_use]
209    pub fn with_tool<T: Tool>(mut self, tool: &T) -> Self {
210        self.tool_definitions.push(tool::ToolDefinition::new(tool));
211        self
212    }
213
214    /// Overrides the sampling parameters used for this call.
215    #[must_use]
216    pub fn with_parameters(mut self, parameters: Parameters) -> Self {
217        self.parameters = parameters;
218        self
219    }
220
221    /// Returns the current conversation messages.
222    #[must_use]
223    pub fn messages(&self) -> &[Message] {
224        &self.messages
225    }
226
227    /// Returns a mutable reference to messages for modification.
228    pub const fn messages_mut(&mut self) -> &mut Vec<Message> {
229        &mut self.messages
230    }
231
232    /// Returns the current parameter snapshot.
233    #[must_use]
234    pub const fn parameters(&self) -> &Parameters {
235        &self.parameters
236    }
237
238    /// Returns the tool definitions.
239    #[must_use]
240    pub fn tool_definitions(&self) -> &[tool::ToolDefinition] {
241        &self.tool_definitions
242    }
243
244    /// Breaks the request into owned components.
245    #[must_use]
246    pub fn into_parts(self) -> (Vec<Message>, Parameters, Vec<tool::ToolDefinition>) {
247        (self.messages, self.parameters, self.tool_definitions)
248    }
249}
250
251/// Legacy request builder that supports mutable tool registry.
252///
253/// This is provided for backwards compatibility with providers that
254/// handle tool execution internally (built-in tools).
255#[derive(Debug)]
256pub struct LLMRequestWithTools<'tools> {
257    inner: LLMRequest,
258    tools: &'tools mut Tools,
259}
260
261impl LLMRequest {
262    /// Attaches a mutable tool registry to the request.
263    ///
264    /// This is for providers that execute tools internally (e.g., built-in tools).
265    /// For agent-controlled tool execution, use `with_tool_definitions` instead.
266    pub fn with_tools(self, tools: &mut Tools) -> LLMRequestWithTools<'_> {
267        let definitions = tools.definitions();
268        LLMRequestWithTools {
269            inner: self.with_tool_definitions(definitions),
270            tools,
271        }
272    }
273}
274
275impl<'tools> LLMRequestWithTools<'tools> {
276    /// Returns the inner request.
277    #[must_use]
278    pub const fn request(&self) -> &LLMRequest {
279        &self.inner
280    }
281
282    /// Returns the tool registry.
283    #[must_use]
284    pub const fn tools(&mut self) -> &mut Tools {
285        self.tools
286    }
287
288    /// Breaks into components.
289    #[must_use]
290    pub fn into_parts(self) -> (LLMRequest, &'tools mut Tools) {
291        (self.inner, self.tools)
292    }
293
294    /// Invokes a registered tool by name, running it with `cx` as its
295    /// [`ToolContext`].
296    ///
297    /// # Errors
298    /// Returns an error if tool is not found or the tool call fails.
299    pub async fn call_tool(
300        &mut self,
301        name: &str,
302        args_json: &str,
303        cx: ToolContext,
304    ) -> crate::Result<ToolResult> {
305        self.tools.call(name, args_json, cx).await
306    }
307}
308
309/// Language models for text generation and conversation.
310///
311/// The `respond` method returns a stream of [`Event`]s. Tool calls are emitted
312/// as events and are NOT automatically executed - this allows higher-level
313/// abstractions (like agents) to control tool execution.
314///
315/// See the [module documentation](crate::llm) for examples and usage patterns.
316pub trait LanguageModel: Sized + Send + Sync {
317    /// The error type returned by this language model.
318    type Error: core::error::Error + Classify + Send + Sync + 'static;
319
320    /// Generates a streaming response to a conversation.
321    ///
322    /// Returns a stream of [`Event`]s including:
323    /// - `Event::Text` - Visible text chunks
324    /// - `Event::Reasoning` - Internal reasoning (for reasoning models)
325    /// - `Event::ToolCall` - Requests to execute tools (NOT auto-executed)
326    /// - `Event::BuiltInToolResult` - Results from provider's built-in tools
327    fn respond(&self, request: LLMRequest)
328    -> impl Stream<Item = Result<Event, Self::Error>> + Send;
329
330    /// Generates a streaming response with a mutable tool registry.
331    ///
332    /// This is for providers that support built-in tool execution.
333    /// The default implementation ignores the tools and delegates to `respond`.
334    fn respond_with_tools(
335        &self,
336        request: LLMRequestWithTools<'_>,
337    ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
338        let (inner, _tools) = request.into_parts();
339        self.respond(inner)
340    }
341
342    /// Generates structured output conforming to JSON schema.
343    ///
344    /// # Note for Implementors
345    /// By default, we use a system prompt to instruct the model to generate structured output based on the provided JSON schema.
346    /// However, that is not efficient enough, if supported, provider should override this method to provide native structured generation support.
347    /// Native structured generation can apply decode rules in token-level, ensuring the output is always valid JSON, and reducing parsing errors.
348    fn generate<T: JsonSchema + DeserializeOwned + 'static>(
349        &self,
350        request: LLMRequest,
351    ) -> impl Future<Output = Result<T, GenerateError<Self::Error>>> + Send {
352        async { structured_generate(self, request).await }
353    }
354
355    /// Completes given text prefix.
356    fn complete(&self, prefix: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
357        self.respond(oneshot("Please complete the following text:", prefix))
358    }
359
360    /// Summarizes text.
361    ///
362    /// # Note for Implementors
363    /// By default, we provide a generic summarization prompt. However, some model provider, like Apple Intelligence, may have native summarization support.
364    /// It would load a summarization-specific lora at runtime, providing better quality.
365    fn summarize(&self, text: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
366        summarize(self, text)
367    }
368
369    /// Categorizes text.
370    ///
371    /// # Note for Implementors
372    /// By default, we use a system prompt to instruct the model to categorize text based on the provided JSON schema.
373    /// However, that is not efficient enough, if supported, provider should override this method to provide native categorization support.
374    fn categorize<T: JsonSchema + DeserializeOwned + 'static>(
375        &self,
376        text: &str,
377    ) -> impl Future<Output = Result<T, GenerateError<Self::Error>>> + Send {
378        async { categorize_text(self, text).await }
379    }
380
381    /// Returns model profile and capabilities.
382    ///
383    /// See [`Profile`] for details on model metadata.
384    ///
385    /// # Errors
386    ///
387    /// Returns an error when the profile cannot be determined — for example
388    /// when neither the provider nor the model registry knows the model's
389    /// context window. Implementations must not substitute a guessed value.
390    fn profile(&self) -> impl Future<Output = Result<Profile, Self::Error>> + Send;
391
392    /// How a request to this model that failed with a retryable failure is
393    /// sent again.
394    ///
395    /// A provider that retries establishing its response stream returns the
396    /// policy it retries under; callers reuse it for a stream that breaks
397    /// after it opened, which the provider cannot retry because the events
398    /// before the break already reached them.
399    fn retry_policy(&self) -> RetryPolicy {
400        RetryPolicy::default()
401    }
402}
403
404macro_rules! impl_language_model {
405    ($($name:ident),*) => {
406        $(
407            impl<T: LanguageModel> LanguageModel for $name<T> {
408                type Error = T::Error;
409
410                fn respond(
411                    &self,
412                    request: LLMRequest,
413                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
414                    T::respond(self, request)
415                }
416
417                fn respond_with_tools(
418                    &self,
419                    request: LLMRequestWithTools<'_>,
420                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
421                    T::respond_with_tools(self, request)
422                }
423
424                fn generate<U: JsonSchema + DeserializeOwned + 'static>(
425                    &self,
426                    request: LLMRequest,
427                ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
428                    T::generate(self, request)
429                }
430
431                fn complete(
432                    &self,
433                    prefix: &str,
434                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
435                    T::complete(self, prefix)
436                }
437
438                fn summarize(
439                    &self,
440                    text: &str,
441                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
442                    T::summarize(self, text)
443                }
444
445                fn categorize<U: JsonSchema + DeserializeOwned + 'static>(
446                    &self,
447                    text: &str,
448                ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
449                    T::categorize(self, text)
450                }
451
452                fn profile(&self) -> impl Future<Output = Result<Profile, Self::Error>> + Send {
453                    T::profile(self)
454                }
455
456                fn retry_policy(&self) -> RetryPolicy {
457                    T::retry_policy(self)
458                }
459            }
460        )*
461    };
462}
463
464impl<T: LanguageModel> LanguageModel for &T {
465    type Error = T::Error;
466
467    fn respond(
468        &self,
469        request: LLMRequest,
470    ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
471        T::respond(self, request)
472    }
473
474    fn respond_with_tools(
475        &self,
476        request: LLMRequestWithTools<'_>,
477    ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
478        T::respond_with_tools(self, request)
479    }
480
481    fn generate<U: JsonSchema + DeserializeOwned + 'static>(
482        &self,
483        request: LLMRequest,
484    ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
485        T::generate(self, request)
486    }
487
488    fn complete(&self, prefix: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
489        T::complete(self, prefix)
490    }
491
492    fn summarize(&self, text: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
493        T::summarize(self, text)
494    }
495
496    fn categorize<U: JsonSchema + DeserializeOwned + 'static>(
497        &self,
498        text: &str,
499    ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
500        T::categorize(self, text)
501    }
502
503    fn profile(&self) -> impl Future<Output = Result<Profile, Self::Error>> + Send {
504        T::profile(self)
505    }
506
507    fn retry_policy(&self) -> RetryPolicy {
508        T::retry_policy(self)
509    }
510}
511
512mod prompts;
513
514impl_language_model!(Arc, Box);
515
516/// Collects text from an event stream.
517///
518/// # Errors
519///
520/// Returns the first stream error encountered while collecting text chunks.
521pub async fn collect_text<S, E>(stream: S) -> Result<String, E>
522where
523    S: Stream<Item = Result<Event, E>>,
524{
525    pin!(stream);
526    let mut result = String::new();
527    while let Some(event) = stream.next().await {
528        if let Event::Text(text) = event? {
529            result.push_str(&text);
530        }
531    }
532    Ok(result)
533}
534
535async fn structured_generate<T: JsonSchema + DeserializeOwned + 'static, M: LanguageModel>(
536    model: &M,
537    mut request: LLMRequest,
538) -> Result<T, GenerateError<M::Error>> {
539    let schema = schema_for!(T);
540
541    // If it is a string, we are not required to set up structured generation and JSON schema.
542    let json = if schema.as_value().is_string() {
543        let stream = model.respond(request);
544        let response = collect_text(stream)
545            .await
546            .map_err(GenerateError::Provider)?;
547        // Let's encode it as JSON string. Serializing a String cannot fail.
548        serde_json::to_string(&response).map_err(|source| GenerateError::Parse {
549            source,
550            response: response.clone(),
551        })?
552    } else {
553        // Serializing a `Schema` is infallible in practice, but keep the error
554        // typed rather than unwrapping.
555        let schema =
556            serde_json::to_string_pretty(&schema).map_err(|source| GenerateError::Parse {
557                source,
558                response: String::new(),
559            })?;
560        let prompt = prompts::generate(&schema);
561        request.messages.push(Message::system(prompt));
562        request.parameters.structured_outputs = true;
563
564        let stream = model.respond(request);
565        collect_text(stream)
566            .await
567            .map_err(GenerateError::Provider)?
568    };
569
570    parse_json_with_recovery(&json).map_err(|source| GenerateError::Parse {
571        source,
572        response: truncate_for_error(&json),
573    })
574}
575
576/// Keeps a failed response short enough to log without dumping a whole context.
577fn truncate_for_error(response: &str) -> String {
578    const LIMIT: usize = 500;
579    response.chars().take(LIMIT).collect()
580}
581
582/// Convenience helper that creates a single system + user [`LLMRequest`].
583pub fn oneshot(system: impl Into<String>, user: impl Into<String>) -> LLMRequest {
584    let messages = vec![Message::system(system.into()), Message::user(user.into())];
585    LLMRequest::new(messages)
586}
587
588fn summarize<M: LanguageModel>(
589    model: &M,
590    text: &str,
591) -> impl Stream<Item = Result<Event, M::Error>> + Send {
592    let messages = oneshot("Summarize text:", text);
593    model.respond(messages)
594}
595
596async fn categorize_text<T: JsonSchema + DeserializeOwned + 'static, M: LanguageModel>(
597    model: &M,
598    text: &str,
599) -> Result<T, GenerateError<M::Error>> {
600    let request = oneshot("Categorize text by provided schema", text);
601    model.generate(request).await
602}
603
604fn parse_json_with_recovery<T: DeserializeOwned + 'static>(
605    json: &str,
606) -> Result<T, serde_json::Error> {
607    use serde::de::Error as _;
608
609    let trimmed = json.trim();
610    let mut last_error: Option<serde_json::Error> = None;
611    let mut last_candidate: Option<String> = None;
612
613    for candidate in build_json_candidates(trimmed) {
614        match serde_json::from_str::<T>(&candidate) {
615            Ok(value) => return Ok(value),
616            Err(err) => {
617                last_error = Some(err);
618                last_candidate = Some(candidate);
619            }
620        }
621    }
622
623    // A model asked for a plain string will often answer with an object or a
624    // bare word instead. Re-encode whatever it did produce as a JSON string.
625    if is_string_type::<T>()
626        && let Some(candidate) = last_candidate
627        && let Ok(value) = serde_json::from_str::<serde_json::Value>(&candidate)
628    {
629        let text = match value {
630            serde_json::Value::String(s) => s,
631            other => other.to_string(),
632        };
633        let encoded = serde_json::to_string(&text)?;
634        if let Ok(value) = serde_json::from_str::<T>(&encoded) {
635            return Ok(value);
636        }
637    }
638
639    Err(last_error.unwrap_or_else(|| {
640        serde_json::Error::custom("structured output was empty or missing a JSON block")
641    }))
642}
643
644fn strip_code_fences(raw: &str) -> Option<String> {
645    let trimmed = raw.trim();
646    let fence_start = trimmed.find("```")?;
647    let after_fence = &trimmed[fence_start + 3..];
648    let mut lines = after_fence.lines();
649    let _maybe_lang = lines.next();
650    let body = lines.collect::<Vec<_>>().join("\n");
651    let content = body.rfind("```").map_or(body.as_str(), |end| &body[..end]);
652
653    let cleaned = content.trim();
654    if cleaned.is_empty() {
655        None
656    } else {
657        Some(cleaned.to_string())
658    }
659}
660
661fn extract_json_block(raw: &str) -> Option<String> {
662    if let (Some(start), Some(end)) = (raw.find('{'), raw.rfind('}'))
663        && end >= start
664    {
665        let candidate = &raw[start..=end];
666        if !candidate.trim().is_empty() {
667            return Some(candidate.trim().to_string());
668        }
669    }
670    if let (Some(start), Some(end)) = (raw.find('['), raw.rfind(']'))
671        && end >= start
672    {
673        let candidate = &raw[start..=end];
674        if !candidate.trim().is_empty() {
675            return Some(candidate.trim().to_string());
676        }
677    }
678    None
679}
680
681fn build_json_candidates(raw: &str) -> Vec<String> {
682    let mut candidates = Vec::new();
683
684    if !raw.is_empty() {
685        candidates.push(raw.to_string());
686    }
687
688    if let Some(fenced) = strip_code_fences(raw) {
689        candidates.push(fenced);
690    }
691
692    if let Some(block) = extract_json_block(raw) {
693        candidates.push(block);
694    }
695
696    if let Some(dequoted) = dequote_json_string(raw) {
697        candidates.push(dequoted);
698    }
699
700    if let Some(stripped) = strip_leading_label(raw, "json") {
701        candidates.push(stripped);
702    }
703
704    let mut deduped = Vec::new();
705    for candidate in candidates {
706        if deduped.iter().all(|seen| seen != &candidate) {
707            deduped.push(candidate);
708        }
709    }
710    deduped
711}
712
713fn dequote_json_string(raw: &str) -> Option<String> {
714    let trimmed = raw.trim();
715    if !(trimmed.starts_with('"') && trimmed.ends_with('"')) {
716        return None;
717    }
718    let inner: String = serde_json::from_str(trimmed).ok()?;
719    if inner.trim().is_empty() {
720        None
721    } else {
722        Some(inner)
723    }
724}
725
726fn strip_leading_label(raw: &str, label: &str) -> Option<String> {
727    let trimmed = raw.trim_start();
728    if !trimmed.to_ascii_lowercase().starts_with(label) {
729        return None;
730    }
731    let stripped = trimmed[label.len()..]
732        .trim_start_matches(|c: char| c.is_whitespace() || c == ':' || c == '-')
733        .trim();
734    if stripped.is_empty() {
735        None
736    } else {
737        Some(stripped.to_string())
738    }
739}
740
741fn is_string_type<T: 'static>() -> bool {
742    TypeId::of::<T>() == TypeId::of::<String>()
743}
744
745#[cfg(test)]
746mod tests {
747    use super::parse_json_with_recovery;
748    use alloc::string::String;
749    use serde::Deserialize;
750
751    #[derive(Debug, Deserialize, PartialEq, Eq)]
752    struct Foo {
753        a: u8,
754    }
755
756    #[test]
757    fn parses_plain_json() {
758        let foo: Foo = parse_json_with_recovery(r#"{"a":1}"#).unwrap();
759        assert_eq!(foo, Foo { a: 1 });
760    }
761
762    #[test]
763    fn parses_code_fence_json() {
764        let foo: Foo = parse_json_with_recovery("```json\n{\"a\":2}\n```").unwrap();
765        assert_eq!(foo, Foo { a: 2 });
766    }
767
768    #[test]
769    fn parses_embedded_block() {
770        let foo: Foo = parse_json_with_recovery("noise {\"a\":3} trailing").unwrap();
771        assert_eq!(foo, Foo { a: 3 });
772    }
773
774    #[test]
775    fn parses_quoted_json_string() {
776        let foo: Foo = parse_json_with_recovery(r#""{\"a\":4}""#).unwrap();
777        assert_eq!(foo, Foo { a: 4 });
778    }
779
780    #[test]
781    fn parses_labeled_json() {
782        let foo: Foo = parse_json_with_recovery("json {\"a\":5}").unwrap();
783        assert_eq!(foo, Foo { a: 5 });
784    }
785
786    #[test]
787    fn coerces_object_to_string() {
788        let value: String =
789            parse_json_with_recovery(r#"{"title":"summary","type":"content"}"#).unwrap();
790        assert!(
791            value.contains("\"title\":\"summary\"") && value.contains("\"type\":\"content\""),
792            "unexpected value: {value}"
793        );
794    }
795}
aither_core/llm/mod.rs

aither_core/llm/
mod.rs
