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 inner request for modification, keeping its tool registry.
283    pub const fn request_mut(&mut self) -> &mut LLMRequest {
284        &mut self.inner
285    }
286
287    /// Returns the tool registry.
288    #[must_use]
289    pub const fn tools(&mut self) -> &mut Tools {
290        self.tools
291    }
292
293    /// Breaks into components.
294    #[must_use]
295    pub fn into_parts(self) -> (LLMRequest, &'tools mut Tools) {
296        (self.inner, self.tools)
297    }
298
299    /// Invokes a registered tool by name, running it with `cx` as its
300    /// [`ToolContext`].
301    ///
302    /// # Errors
303    /// Returns an error if tool is not found or the tool call fails.
304    pub async fn call_tool(
305        &mut self,
306        name: &str,
307        args_json: &str,
308        cx: ToolContext,
309    ) -> crate::Result<ToolResult> {
310        self.tools.call(name, args_json, cx).await
311    }
312}
313
314/// Language models for text generation and conversation.
315///
316/// The `respond` method returns a stream of [`Event`]s. Tool calls are emitted
317/// as events and are NOT automatically executed - this allows higher-level
318/// abstractions (like agents) to control tool execution.
319///
320/// See the [module documentation](crate::llm) for examples and usage patterns.
321pub trait LanguageModel: Sized + Send + Sync {
322    /// The error type returned by this language model.
323    type Error: core::error::Error + Classify + Send + Sync + 'static;
324
325    /// Generates a streaming response to a conversation.
326    ///
327    /// Returns a stream of [`Event`]s including:
328    /// - `Event::Text` - Visible text chunks
329    /// - `Event::Reasoning` - Internal reasoning (for reasoning models)
330    /// - `Event::ToolCall` - Requests to execute tools (NOT auto-executed)
331    /// - `Event::BuiltInToolResult` - Results from provider's built-in tools
332    fn respond(&self, request: LLMRequest)
333    -> impl Stream<Item = Result<Event, Self::Error>> + Send;
334
335    /// Generates a streaming response with a mutable tool registry.
336    ///
337    /// This is for providers that support built-in tool execution.
338    /// The default implementation ignores the tools and delegates to `respond`.
339    fn respond_with_tools(
340        &self,
341        request: LLMRequestWithTools<'_>,
342    ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
343        let (inner, _tools) = request.into_parts();
344        self.respond(inner)
345    }
346
347    /// Generates structured output conforming to JSON schema.
348    ///
349    /// # Note for Implementors
350    /// By default, we use a system prompt to instruct the model to generate structured output based on the provided JSON schema.
351    /// However, that is not efficient enough, if supported, provider should override this method to provide native structured generation support.
352    /// Native structured generation can apply decode rules in token-level, ensuring the output is always valid JSON, and reducing parsing errors.
353    fn generate<T: JsonSchema + DeserializeOwned + 'static>(
354        &self,
355        request: LLMRequest,
356    ) -> impl Future<Output = Result<T, GenerateError<Self::Error>>> + Send {
357        async { structured_generate(self, request).await }
358    }
359
360    /// Completes given text prefix.
361    fn complete(&self, prefix: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
362        self.respond(oneshot("Please complete the following text:", prefix))
363    }
364
365    /// Summarizes text.
366    ///
367    /// # Note for Implementors
368    /// By default, we provide a generic summarization prompt. However, some model provider, like Apple Intelligence, may have native summarization support.
369    /// It would load a summarization-specific lora at runtime, providing better quality.
370    fn summarize(&self, text: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
371        summarize(self, text)
372    }
373
374    /// Categorizes text.
375    ///
376    /// # Note for Implementors
377    /// By default, we use a system prompt to instruct the model to categorize text based on the provided JSON schema.
378    /// However, that is not efficient enough, if supported, provider should override this method to provide native categorization support.
379    fn categorize<T: JsonSchema + DeserializeOwned + 'static>(
380        &self,
381        text: &str,
382    ) -> impl Future<Output = Result<T, GenerateError<Self::Error>>> + Send {
383        async { categorize_text(self, text).await }
384    }
385
386    /// Returns model profile and capabilities.
387    ///
388    /// See [`Profile`] for details on model metadata.
389    ///
390    /// # Errors
391    ///
392    /// Returns an error when the profile cannot be determined — for example
393    /// when neither the provider nor the model registry knows the model's
394    /// context window. Implementations must not substitute a guessed value.
395    fn profile(&self) -> impl Future<Output = Result<Profile, Self::Error>> + Send;
396
397    /// How a request to this model that failed with a retryable failure is
398    /// sent again.
399    ///
400    /// A provider that retries establishing its response stream returns the
401    /// policy it retries under; callers reuse it for a stream that breaks
402    /// after it opened, which the provider cannot retry because the events
403    /// before the break already reached them.
404    fn retry_policy(&self) -> RetryPolicy {
405        RetryPolicy::default()
406    }
407}
408
409macro_rules! impl_language_model {
410    ($($name:ident),*) => {
411        $(
412            impl<T: LanguageModel> LanguageModel for $name<T> {
413                type Error = T::Error;
414
415                fn respond(
416                    &self,
417                    request: LLMRequest,
418                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
419                    T::respond(self, request)
420                }
421
422                fn respond_with_tools(
423                    &self,
424                    request: LLMRequestWithTools<'_>,
425                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
426                    T::respond_with_tools(self, request)
427                }
428
429                fn generate<U: JsonSchema + DeserializeOwned + 'static>(
430                    &self,
431                    request: LLMRequest,
432                ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
433                    T::generate(self, request)
434                }
435
436                fn complete(
437                    &self,
438                    prefix: &str,
439                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
440                    T::complete(self, prefix)
441                }
442
443                fn summarize(
444                    &self,
445                    text: &str,
446                ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
447                    T::summarize(self, text)
448                }
449
450                fn categorize<U: JsonSchema + DeserializeOwned + 'static>(
451                    &self,
452                    text: &str,
453                ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
454                    T::categorize(self, text)
455                }
456
457                fn profile(&self) -> impl Future<Output = Result<Profile, Self::Error>> + Send {
458                    T::profile(self)
459                }
460
461                fn retry_policy(&self) -> RetryPolicy {
462                    T::retry_policy(self)
463                }
464            }
465        )*
466    };
467}
468
469impl<T: LanguageModel> LanguageModel for &T {
470    type Error = T::Error;
471
472    fn respond(
473        &self,
474        request: LLMRequest,
475    ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
476        T::respond(self, request)
477    }
478
479    fn respond_with_tools(
480        &self,
481        request: LLMRequestWithTools<'_>,
482    ) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
483        T::respond_with_tools(self, request)
484    }
485
486    fn generate<U: JsonSchema + DeserializeOwned + 'static>(
487        &self,
488        request: LLMRequest,
489    ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
490        T::generate(self, request)
491    }
492
493    fn complete(&self, prefix: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
494        T::complete(self, prefix)
495    }
496
497    fn summarize(&self, text: &str) -> impl Stream<Item = Result<Event, Self::Error>> + Send {
498        T::summarize(self, text)
499    }
500
501    fn categorize<U: JsonSchema + DeserializeOwned + 'static>(
502        &self,
503        text: &str,
504    ) -> impl Future<Output = Result<U, GenerateError<Self::Error>>> + Send {
505        T::categorize(self, text)
506    }
507
508    fn profile(&self) -> impl Future<Output = Result<Profile, Self::Error>> + Send {
509        T::profile(self)
510    }
511
512    fn retry_policy(&self) -> RetryPolicy {
513        T::retry_policy(self)
514    }
515}
516
517mod prompts;
518
519impl_language_model!(Arc, Box);
520
521/// Collects text from an event stream.
522///
523/// # Errors
524///
525/// Returns the first stream error encountered while collecting text chunks.
526pub async fn collect_text<S, E>(stream: S) -> Result<String, E>
527where
528    S: Stream<Item = Result<Event, E>>,
529{
530    pin!(stream);
531    let mut result = String::new();
532    while let Some(event) = stream.next().await {
533        if let Event::Text(text) = event? {
534            result.push_str(&text);
535        }
536    }
537    Ok(result)
538}
539
540async fn structured_generate<T: JsonSchema + DeserializeOwned + 'static, M: LanguageModel>(
541    model: &M,
542    mut request: LLMRequest,
543) -> Result<T, GenerateError<M::Error>> {
544    let schema = schema_for!(T);
545
546    // If it is a string, we are not required to set up structured generation and JSON schema.
547    let json = if schema.as_value().is_string() {
548        let stream = model.respond(request);
549        let response = collect_text(stream)
550            .await
551            .map_err(GenerateError::Provider)?;
552        // Let's encode it as JSON string. Serializing a String cannot fail.
553        serde_json::to_string(&response).map_err(|source| GenerateError::Parse {
554            source,
555            response: response.clone(),
556        })?
557    } else {
558        // Serializing a `Schema` is infallible in practice, but keep the error
559        // typed rather than unwrapping.
560        let schema =
561            serde_json::to_string_pretty(&schema).map_err(|source| GenerateError::Parse {
562                source,
563                response: String::new(),
564            })?;
565        let prompt = prompts::generate(&schema);
566        request.messages.push(Message::system(prompt));
567        request.parameters.structured_outputs = true;
568
569        let stream = model.respond(request);
570        collect_text(stream)
571            .await
572            .map_err(GenerateError::Provider)?
573    };
574
575    parse_json_with_recovery(&json).map_err(|source| GenerateError::Parse {
576        source,
577        response: truncate_for_error(&json),
578    })
579}
580
581/// Keeps a failed response short enough to log without dumping a whole context.
582fn truncate_for_error(response: &str) -> String {
583    const LIMIT: usize = 500;
584    response.chars().take(LIMIT).collect()
585}
586
587/// Convenience helper that creates a single system + user [`LLMRequest`].
588pub fn oneshot(system: impl Into<String>, user: impl Into<String>) -> LLMRequest {
589    let messages = vec![Message::system(system.into()), Message::user(user.into())];
590    LLMRequest::new(messages)
591}
592
593fn summarize<M: LanguageModel>(
594    model: &M,
595    text: &str,
596) -> impl Stream<Item = Result<Event, M::Error>> + Send {
597    let messages = oneshot("Summarize text:", text);
598    model.respond(messages)
599}
600
601async fn categorize_text<T: JsonSchema + DeserializeOwned + 'static, M: LanguageModel>(
602    model: &M,
603    text: &str,
604) -> Result<T, GenerateError<M::Error>> {
605    let request = oneshot("Categorize text by provided schema", text);
606    model.generate(request).await
607}
608
609fn parse_json_with_recovery<T: DeserializeOwned + 'static>(
610    json: &str,
611) -> Result<T, serde_json::Error> {
612    use serde::de::Error as _;
613
614    let trimmed = json.trim();
615    let mut last_error: Option<serde_json::Error> = None;
616    let mut last_candidate: Option<String> = None;
617
618    for candidate in build_json_candidates(trimmed) {
619        match serde_json::from_str::<T>(&candidate) {
620            Ok(value) => return Ok(value),
621            Err(err) => {
622                last_error = Some(err);
623                last_candidate = Some(candidate);
624            }
625        }
626    }
627
628    // A model asked for a plain string will often answer with an object or a
629    // bare word instead. Re-encode whatever it did produce as a JSON string.
630    if is_string_type::<T>()
631        && let Some(candidate) = last_candidate
632        && let Ok(value) = serde_json::from_str::<serde_json::Value>(&candidate)
633    {
634        let text = match value {
635            serde_json::Value::String(s) => s,
636            other => other.to_string(),
637        };
638        let encoded = serde_json::to_string(&text)?;
639        if let Ok(value) = serde_json::from_str::<T>(&encoded) {
640            return Ok(value);
641        }
642    }
643
644    Err(last_error.unwrap_or_else(|| {
645        serde_json::Error::custom("structured output was empty or missing a JSON block")
646    }))
647}
648
649fn strip_code_fences(raw: &str) -> Option<String> {
650    let trimmed = raw.trim();
651    let fence_start = trimmed.find("```")?;
652    let after_fence = &trimmed[fence_start + 3..];
653    let mut lines = after_fence.lines();
654    let _maybe_lang = lines.next();
655    let body = lines.collect::<Vec<_>>().join("\n");
656    let content = body.rfind("```").map_or(body.as_str(), |end| &body[..end]);
657
658    let cleaned = content.trim();
659    if cleaned.is_empty() {
660        None
661    } else {
662        Some(cleaned.to_string())
663    }
664}
665
666fn extract_json_block(raw: &str) -> Option<String> {
667    if let (Some(start), Some(end)) = (raw.find('{'), raw.rfind('}'))
668        && end >= start
669    {
670        let candidate = &raw[start..=end];
671        if !candidate.trim().is_empty() {
672            return Some(candidate.trim().to_string());
673        }
674    }
675    if let (Some(start), Some(end)) = (raw.find('['), raw.rfind(']'))
676        && end >= start
677    {
678        let candidate = &raw[start..=end];
679        if !candidate.trim().is_empty() {
680            return Some(candidate.trim().to_string());
681        }
682    }
683    None
684}
685
686fn build_json_candidates(raw: &str) -> Vec<String> {
687    let mut candidates = Vec::new();
688
689    if !raw.is_empty() {
690        candidates.push(raw.to_string());
691    }
692
693    if let Some(fenced) = strip_code_fences(raw) {
694        candidates.push(fenced);
695    }
696
697    if let Some(block) = extract_json_block(raw) {
698        candidates.push(block);
699    }
700
701    if let Some(dequoted) = dequote_json_string(raw) {
702        candidates.push(dequoted);
703    }
704
705    if let Some(stripped) = strip_leading_label(raw, "json") {
706        candidates.push(stripped);
707    }
708
709    let mut deduped = Vec::new();
710    for candidate in candidates {
711        if deduped.iter().all(|seen| seen != &candidate) {
712            deduped.push(candidate);
713        }
714    }
715    deduped
716}
717
718fn dequote_json_string(raw: &str) -> Option<String> {
719    let trimmed = raw.trim();
720    if !(trimmed.starts_with('"') && trimmed.ends_with('"')) {
721        return None;
722    }
723    let inner: String = serde_json::from_str(trimmed).ok()?;
724    if inner.trim().is_empty() {
725        None
726    } else {
727        Some(inner)
728    }
729}
730
731fn strip_leading_label(raw: &str, label: &str) -> Option<String> {
732    let trimmed = raw.trim_start();
733    if !trimmed.to_ascii_lowercase().starts_with(label) {
734        return None;
735    }
736    let stripped = trimmed[label.len()..]
737        .trim_start_matches(|c: char| c.is_whitespace() || c == ':' || c == '-')
738        .trim();
739    if stripped.is_empty() {
740        None
741    } else {
742        Some(stripped.to_string())
743    }
744}
745
746fn is_string_type<T: 'static>() -> bool {
747    TypeId::of::<T>() == TypeId::of::<String>()
748}
749
750#[cfg(test)]
751mod tests {
752    use super::parse_json_with_recovery;
753    use alloc::string::String;
754    use serde::Deserialize;
755
756    #[derive(Debug, Deserialize, PartialEq, Eq)]
757    struct Foo {
758        a: u8,
759    }
760
761    #[test]
762    fn parses_plain_json() {
763        let foo: Foo = parse_json_with_recovery(r#"{"a":1}"#).unwrap();
764        assert_eq!(foo, Foo { a: 1 });
765    }
766
767    #[test]
768    fn parses_code_fence_json() {
769        let foo: Foo = parse_json_with_recovery("```json\n{\"a\":2}\n```").unwrap();
770        assert_eq!(foo, Foo { a: 2 });
771    }
772
773    #[test]
774    fn parses_embedded_block() {
775        let foo: Foo = parse_json_with_recovery("noise {\"a\":3} trailing").unwrap();
776        assert_eq!(foo, Foo { a: 3 });
777    }
778
779    #[test]
780    fn parses_quoted_json_string() {
781        let foo: Foo = parse_json_with_recovery(r#""{\"a\":4}""#).unwrap();
782        assert_eq!(foo, Foo { a: 4 });
783    }
784
785    #[test]
786    fn parses_labeled_json() {
787        let foo: Foo = parse_json_with_recovery("json {\"a\":5}").unwrap();
788        assert_eq!(foo, Foo { a: 5 });
789    }
790
791    #[test]
792    fn coerces_object_to_string() {
793        let value: String =
794            parse_json_with_recovery(r#"{"title":"summary","type":"content"}"#).unwrap();
795        assert!(
796            value.contains("\"title\":\"summary\"") && value.contains("\"type\":\"content\""),
797            "unexpected value: {value}"
798        );
799    }
800}
aither_core/llm/mod.rs

aither_core/llm/
mod.rs
