OllamaDiscovery

Category: Providers

Source: ollama_discovery.dart

Read-only client for Ollama's metadata endpoints.

Today we only need /api/tags — the list of locally-pulled models — so the model picker can stop lying about which catalog entries are actually runnable. The client is fail-soft by construction: every network error, timeout, non-200, or malformed JSON collapses to an empty list. Callers never need to try/catch.

A tiny in-memory cache (30 s TTL keyed by base URL) absorbs the case where a user opens the picker repeatedly within a short window.

Classes

OllamaInstalledModel

One entry from GET /api/tags — the fields we actually use.

Constructor

const OllamaInstalledModel({
    required this.tag,
    required this.sizeBytes,
    required this.modifiedAt,
  })

Properties

Property	Type	Description
tag	String	Fully-qualified tag as Ollama reports it, e.g. qwen3-coder:30b.
sizeBytes	int
modifiedAt	DateTime?

OllamaPullProgress

One NDJSON frame from POST /api/pull. Ollama emits a status string (pulling manifest, downloading, success, …) and, for download frames, total/completed byte counters.

Constructor

const OllamaPullProgress({
    required this.status,
    this.total,
    this.completed,
    this.digest,
    this.error,
  })

Properties

Property	Type	Description
status	String
total	int?
completed	int?
digest	String?
error	String?
isSuccess	bool get
hasError	bool get
isSuccess	bool get
hasError	bool get
fraction	double? get	0.0..1.0 when download byte counters are present, else null.

OllamaDiscovery

Constructor

OllamaDiscovery({
    required this.baseUrl,
    http.Client Function()? clientFactory,
    this.timeout = const Duration(seconds: 2),
    this.cacheTtl = const Duration(seconds: 30),
    DateTime Function()? now,
  })

Properties

Property	Type	Description
baseUrl	Uri	Ollama's HTTP root. Accepts the /v1 suffix (OpenAI-compat path) and the bare form — we normalise internally so callers don't have to.
timeout	Duration
cacheTtl	Duration
models	List<OllamaInstalledModel>
expiresAt	DateTime

Methods

Future<List<OllamaInstalledModel>> listInstalled()

GET /api/tags. Returns an empty list on any failure — never throws.

void invalidateCache()

Drop the cache entry for this base URL. Tests and /model refresh call this; normal picker open should rely on the TTL.

Stream<OllamaPullProgress> pullModel(String tag)

Stream NDJSON progress from POST /api/pull. The stream completes when Ollama finishes (a {"status":"success"} frame) or errors out. On error the final frame carries [OllamaPullProgress.error] so callers can surface it without try/catch boilerplate.

static void resetCacheForTesting()

For tests: wipe all cached entries regardless of base URL.