Struct PaddockProvider 

pub struct PaddockProvider { /* private fields */ }

Implementations

impl PaddockProvider

pub fn new( base_url: String, model: String, api_key: Option<String>, ) -> PaddockProvider

Trait Implementations

impl LlmProvider for PaddockProvider

fn chat_json<'life0, 'life1, 'life2, 'life3, 'life4, 'async_trait>( &'life0 self, system: &'life1 str, msgs: &'life2 [Msg], schema: &'life3 Value, name: &'life4 str, ) -> Pin<Box<dyn Future<Output = Result<Option<Value>, String>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait, 'life4: 'async_trait,

Constrain decoding to schema via the OpenAI-compatible response_format.

Ollama, llama.cpp’s server, vLLM and LM Studio all accept this and compile the schema into a decoding grammar internally, so the model can only emit tokens the schema permits. That is why this works on models with no tool-calling ability at all: nothing is being asked of the model except to continue, and the sampler does the rest.

Two degradations, because not every server implements the whole thing:

• a server that rejects json_schema is retried with json_object, which constrains the output to some JSON and leaves the shape to the prompt;
• a server that rejects response_format entirely returns Ok(None), so the caller falls back rather than seeing an error it cannot act on.

fn name(&self) -> &str

fn chat<'life0, 'life1, 'life2, 'life3, 'async_trait>( &'life0 self, system: &'life1 str, msgs: &'life2 [Msg], tools: &'life3 [ToolSpec], ) -> Pin<Box<dyn Future<Output = Result<Turn, String>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait,

One turn: given the system prompt, conversation, and available tools, return the model’s reply (assistant text and/or tool calls).

fn synthesizes(&self) -> bool

Whether this provider produces trustworthy natural-language synthesis of tool results. Large models (Bedrock, big OpenAI-compatible) do; small on-device tool-callers (Needle, the native 1.7B) do NOT — for those the agent renders a deterministic template from the program result instead of trusting the model to phrase the numbers. Default true.