@ondeinference / onde
Code Commits Issues Pulls Cloud Agent CI/CD
development
copilot/assess-inference-engine-complexity copilot/fix-onde-npm-sdk-v1-1-2-release development feature/diffusion-vision main
ts 137 lines 5.14 KB
Raw
1 // ── Message role ─────────────────────────────────────────────────────────────
2
3 /** Role of a message in a chat conversation. */
4 export type ChatRole = "system" | "user" | "assistant";
5
6 // ── Chat message ─────────────────────────────────────────────────────────────
7
8 /** A single message in a conversation. */
9 export interface ChatMessage {
10 role: ChatRole;
11 content: string;
12 }
13
14 // ── Sampling configuration ───────────────────────────────────────────────────
15
16 /**
17 * Sampling parameters for text generation.
18 * All fields are optional — `undefined` means "use the engine default".
19 */
20 export interface SamplingConfig {
21 /** Sampling temperature (higher = more random). Typical range: 0.0–2.0. */
22 temperature?: number;
23 /** Nucleus (top-p) sampling threshold. Typical value: 0.9–0.95. */
24 topP?: number;
25 /** Top-k sampling limit. */
26 topK?: number;
27 /** Min-p sampling threshold. */
28 minP?: number;
29 /** Maximum number of tokens to generate. */
30 maxTokens?: number;
31 /** Frequency penalty (penalise tokens proportional to occurrence count). */
32 frequencyPenalty?: number;
33 /** Presence penalty (penalise tokens that appeared at all). */
34 presencePenalty?: number;
35 }
36
37 // ── Model configuration ──────────────────────────────────────────────────────
38
39 /** Configuration for loading a GGUF model. */
40 export interface GgufModelConfig {
41 /** HuggingFace model repository ID. */
42 modelId: string;
43 /** GGUF filename(s) within the repository. */
44 files: string[];
45 /** Optional: explicit tokenizer model ID (required on Android). */
46 tokModelId?: string;
47 /** Human-friendly display name. */
48 displayName: string;
49 /** Approximate memory footprint description. */
50 approxMemory: string;
51 /** Optional Jinja chat template override. */
52 chatTemplate?: string;
53 }
54
55 /** Configuration for loading a UQFF model. */
56 export interface UqffModelConfig {
57 /** HuggingFace base model repository ID or local model directory. */
58 modelId: string;
59 /** UQFF shard filename(s), first shard, or shorthand such as "q4k" / "4". */
60 files: string[];
61 /** Human-friendly display name. */
62 displayName: string;
63 /** Approximate memory footprint description. */
64 approxMemory: string;
65 /** Optional Jinja chat template override. */
66 chatTemplate?: string;
67 }
68
69 // ── Inference result ─────────────────────────────────────────────────────────
70
71 /** Result of a completed inference request. */
72 export interface InferenceResult {
73 /** The generated text. */
74 text: string;
75 /** Wall-clock inference duration in seconds. */
76 durationSecs: number;
77 /** Human-readable duration string, e.g. "1.23s". */
78 durationDisplay: string;
79 /** Finish reason, if available. */
80 finishReason?: string;
81 /** Tool calls requested by the model, if any. */
82 toolCalls: ToolCallInfo[];
83 }
84
85 /** Metadata about a tool call requested by the model. */
86 export interface ToolCallInfo {
87 /** Unique identifier for this tool call. */
88 id: string;
89 /** Name of the function to call. */
90 functionName: string;
91 /** JSON-encoded arguments for the function. */
92 arguments: string;
93 }
94
95 // ── Streaming ────────────────────────────────────────────────────────────────
96
97 /** A single streaming token chunk. */
98 export interface StreamChunk {
99 /** The token text delta. */
100 delta: string;
101 /** Whether this is the final chunk. */
102 done: boolean;
103 /** Finish reason (only present on the final chunk). */
104 finishReason?: string;
105 }
106
107 // ── Engine status ────────────────────────────────────────────────────────────
108
109 /** Current status of the inference engine. */
110 export type EngineStatus =
111 | "Unloaded"
112 | "Loading"
113 | "Ready"
114 | "Generating"
115 | "Error";
116
117 /** Snapshot of the engine's current state. */
118 export interface EngineInfo {
119 /** Current engine status. */
120 status: EngineStatus;
121 /** Name of the loaded model, if any. */
122 modelName?: string;
123 /** Approximate memory footprint of the loaded model. */
124 approxMemory?: string;
125 /** Number of messages in the conversation history. */
126 historyLength: number;
127 }
128
129 // ── Errors ───────────────────────────────────────────────────────────────────
130
131 /** Error thrown by the Onde inference engine. */
132 export class OndeError extends Error {
133 constructor(message: string) {
134 super(message);
135 this.name = "OndeError";
136 }
137 }