artificial_core/provider/
chat_complete.rs

1use std::{future::Future, pin::Pin};
2
3use crate::{
4    error::Result,
5    generic::{GenericChatCompletionResponse, GenericFunctionSpec, GenericMessage},
6    model::Model,
7};
8use futures_core::stream::Stream;
9
10/// A **backend** turns a chat prompt into a network call to a concrete provider
11/// (OpenAI, Ollama, Anthropic, …) and parses the structured chat response.
12///
13/// The trait is intentionally minimal:
14///
15/// * **One associated type** – the in-memory `Message` representation this
16///   provider accepts.
17/// * **One async-ish method** – `chat_complete`, which performs a *single*
18///   non-streaming round-trip and returns a value whose type is dictated by
19///   the `PromptTemplate`.
20pub trait ChatCompletionProvider: Send + Sync {
21    /// Chat message type consumed by this backend.
22    type Message: Send + Sync + 'static;
23
24    /// Execute the chat prompt and deserialize the provider’s reply into
25    /// `P::Output`.
26    fn chat_complete<'p, M>(
27        &self,
28        params: ChatCompleteParameters<M>,
29    ) -> Pin<
30        Box<dyn Future<Output = Result<GenericChatCompletionResponse<GenericMessage>>> + Send + 'p>,
31    >
32    where
33        M: Into<Self::Message> + Send + Sync + 'p;
34}
35
36/// A provider that can deliver the model’s answer **incrementally**.
37///
38/// The stream yields UTF-8 text *deltas* (similar to OpenAI’s SSE format).
39/// Tool-call and richer payload support can be layered on later by
40/// introducing a dedicated enum – starting with plain text keeps the API
41/// minimal and backend-agnostic.
42pub trait StreamingChatProvider: ChatCompletionProvider {
43    /// The item type returned on the stream.  For now it is plain UTF-8 text
44    /// chunks, but back-ends are free to wrap it in richer enums if needed.
45    type Delta<'s>: Stream<Item = Result<String>> + Send + 's
46    where
47        Self: 's;
48
49    /// Start a streaming chat completion.
50    fn chat_complete_stream<'p, M>(&self, params: ChatCompleteParameters<M>) -> Self::Delta<'p>
51    where
52        M: Into<Self::Message> + Send + Sync + 'p;
53}
54
55pub struct ChatCompleteParameters<M> {
56    pub messages: Vec<M>,
57    pub model: Model,
58    pub tools: Option<Vec<GenericFunctionSpec>>,
59    pub temperature: Option<f64>,
60    pub response_format: Option<serde_json::Value>,
61}
62
63impl<M> ChatCompleteParameters<M> {
64    pub fn new(messages: Vec<M>, model: Model) -> Self {
65        Self {
66            messages,
67            model,
68            tools: None,
69            temperature: None,
70            response_format: None,
71        }
72    }
73
74    pub fn messages(&self) -> &Vec<M> {
75        &self.messages
76    }
77
78    pub fn model(&self) -> Model {
79        self.model.clone()
80    }
81
82    pub fn tools(&self) -> Option<&Vec<GenericFunctionSpec>> {
83        self.tools.as_ref()
84    }
85
86    pub fn with_temperature(mut self, temperature: f64) -> Self {
87        self.temperature = Some(temperature);
88        self
89    }
90
91    pub fn with_response_format(mut self, response_format: serde_json::Value) -> Self {
92        self.response_format = Some(response_format);
93        self
94    }
95
96    pub fn with_tools(mut self, tools: Vec<GenericFunctionSpec>) -> Self {
97        self.tools = Some(tools);
98        self
99    }
100}
artificial_core/provider/chat_complete.rs

artificial_core/provider/
chat_complete.rs
