class

LLM::Ollama

Ollama LLM client with context-aware request support

Constants

JSON_MODE = JSON::Any.new("json")

format: "json" — Ollama's plain JSON mode.

TEMPERATURE = 0.3

Endpoint extraction wants the model to read code, not to write prose about it, so both request paths pin a low temperature.

Constructors

new(url : String, model : String)

Source

Class methods

format_value(format : String) : JSON::Any

Ollama's format field takes either the literal string "json" or a raw JSON Schema, and uses a schema to constrain decoding. The formats in LLM::* are OpenAI-shaped envelopes ({"type":"json_schema","json_schema":{"schema":{...}}}); handing that envelope straight to Ollama constrained generation to the envelope rather than to the endpoint object we asked for, so the response never matched what the analyzer parses — every endpoint in the request was lost. Unwrap to the inner schema, and fall back to plain JSON mode for anything we don't recognise.

Source

Instance methods

request(prompt : String, format : String = "json")

Make a simple request without context management

Source

request_with_context(system : String | Nil, user : String, format : String = "json", cache_key : String | Nil = nil)

Make a request with optional context management for improved efficiency

Source