Press n or j to go to the next uncovered block, b, p or k for the previous block.
| 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 | 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 4703x 4703x 16144x 16144x 258223x 258223x 16144x 4703x 4703x 4703x 4703x 4703x 4703x 4703x 1x 1x 1x 1x 139x 139x 1x 4647x 4647x 1x 1x 1x 1x 7x 1x 1x 1x 1x 7x 7x 7x 7x 7x 7x 7x 7x 7x 1x 8x 8x 8x 8x 8x 8x 8x 8x 8x 1x 1x 7x 8x 8x 2x 2x 5x 5x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 283x 268x 268x 32x 4x 4x 17x 10x 10x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 3x 3x 1x 1x 3x 3x 3x 1x 1x 3x 1x 1x 2x 2x 2x 1x 1x 1x 1x 1x 1x 1x 1x 1x 2x 2x 2x 1x 1x 1x 1x 1x 1x 1x 3x 2x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 6x 6x 3x 3x 6x 1x 1x 2x 2x 1x 1x 1x 1x 1x 1x 1x 1x 152x 152x 139x 139x 16x 9x 2x 2x 7x 7x 7x 7x 7x 9x 9x 9x 9x 9x 3x 3x 15x 152x 152x | // cspell:ignore huggingface Xenova onnx
/**
* Embedding providers (PRD R5).
*
* Three implementations behind one interface: a deterministic offline provider
* (the default — no keys, no network, no model download, so tests and CI are
* hermetic), an ONNX-backed local model, and a remote OpenAI-compatible
* endpoint. Switching between them changes no consumer; `createProvider`
* selects one from the configuration `app.ts` takes as an option.
*
* This is executed code, not a declaration module. It is also the module the
* offline audit declares as the service's network exception (see
* `test/offline.test.ts`): the remote provider and the local model's first-use
* download are the only places this package reaches the network.
*/
import {createHash} from 'node:crypto';
export interface EmbeddingProvider {
readonly dimension: number;
embed(text: string): Promise<number[]>;
}
/**
* The slice of `@huggingface/transformers` this package uses. Structural, so
* the optional package never has to be installed for the package to
* type-check, and so a test can stand in for it.
*/
export interface TransformersModule {
pipeline: (
task: string,
model: string,
) => Promise<
(
input: string,
options: {pooling: string; normalize: boolean},
) => Promise<{data: Float32Array}>
>;
}
export interface EmbeddingConfig {
kind: 'offline' | 'local' | 'remote';
/** Offline dimensionality. */
dimension?: number;
/** Remote endpoint and model. */
url?: string;
model?: string;
apiKey?: string;
/** Injected for tests. */
fetch?: typeof fetch;
/**
* Loads the optional `@huggingface/transformers` package. Injected so the
* local provider is testable without the heavyweight dependency (and
* without the model it would download on first use); the default performs
* the real dynamic import.
*/
loadTransformers?: () => Promise<TransformersModule>;
}
/**
* Deterministic unit-norm vector derived from the text. Two identical
* signatures always yield the same vector and a dot product is therefore a
* cosine similarity — enough for structural clustering without a model.
*/
export function hashEmbedding(text: string, dimension: number): number[] {
const values: number[] = [];
for (let block = 0; values.length < dimension; block++) {
const digest = createHash('sha256').update(`${block}:${text}`).digest();
for (let i = 0; i < digest.length && values.length < dimension; i += 2) {
values.push(digest.readUInt16BE(i) / 32767.5 - 1);
}
}
// Every component is derived from a 16-bit integer, so none of them is zero
// (32767.5 is not an integer) and the norm is positive for any width above
// zero. A `|| 1` fallback here would be an unreachable branch hiding an
// untested guard, which is why the normalisation has none.
const norm = Math.sqrt(values.reduce((sum, value) => sum + value * value, 0));
return values.map(value => value / norm);
}
class OfflineProvider implements EmbeddingProvider {
readonly dimension: number;
constructor(dimension: number) {
this.dimension = dimension;
}
async embed(text: string): Promise<number[]> {
return hashEmbedding(text, this.dimension);
}
}
class RemoteProvider implements EmbeddingProvider {
readonly dimension: number;
private readonly config: Required<Pick<EmbeddingConfig, 'url' | 'model'>> & {
apiKey?: string;
fetch: typeof fetch;
};
constructor(
config: Required<Pick<EmbeddingConfig, 'url' | 'model'>> & {
apiKey?: string;
fetch: typeof fetch;
},
dimension: number,
) {
this.config = config;
this.dimension = dimension;
}
async embed(text: string): Promise<number[]> {
const response = await this.config.fetch(this.config.url, {
method: 'POST',
headers: {
'content-type': 'application/json',
...(this.config.apiKey ? {authorization: `Bearer ${this.config.apiKey}`} : {}),
},
body: JSON.stringify({input: text, model: this.config.model}),
});
if (!response.ok) {
throw new Error(`embedding provider failed: ${response.status} ${response.statusText}`);
}
const body = (await response.json()) as {data?: Array<{embedding?: number[]}>};
const embedding = body.data?.[0]?.embedding;
if (!embedding) {
throw new Error('embedding provider returned no vector');
}
return embedding;
}
}
/** The optional package the local provider loads, named here so errors can name it too. */
const LOCAL_PACKAGE = '@huggingface/transformers';
/** The ONNX model the local provider downloads on first use — the network is reached through the dependency. */
export const LOCAL_MODEL = 'Xenova/all-MiniLM-L6-v2';
/**
* The model the offline provider is, named for the identity below.
*
* It has no model in any machine-learning sense — it is a seeded hash — and saying so
* is the point: a snapshot written under it must not be restored as if its vectors came
* from a real embedding space, because they cannot answer a natural-language query.
*/
export const OFFLINE_MODEL = 'sha256-hash';
/**
* Which provider a set of vectors came from (D26).
*
* A vector is only comparable with a vector from the same provider at the same width:
* the same text embedded by `sha256-hash` and by `all-MiniLM-L6-v2` share no geometry at
* all, and two widths of one provider do not even share a dimension. So the identity
* travels with the vectors, and a snapshot written under one must not be restored under
* another — the numbers would still add up and every answer would be nonsense.
*/
export interface ProviderIdentity {
kind: 'offline' | 'local' | 'remote';
model: string;
dimension: number;
}
/**
* The identity of the provider a configuration selects.
*
* The dimension defaults live here rather than in `createProvider`, so the identity a
* snapshot is written under and the provider that actually embeds cannot disagree about
* how wide the vectors are.
*
* A remote configuration with no `model` yields `''` here and is rejected by
* `createProvider`: the identity exists to be written into a snapshot, and a
* configuration that cannot build a provider never reaches one.
*/
export function providerIdentity(config: EmbeddingConfig): ProviderIdentity {
if (config.kind === 'offline') {
return {kind: 'offline', model: OFFLINE_MODEL, dimension: config.dimension ?? 64};
}
if (config.kind === 'local') {
return {kind: 'local', model: LOCAL_MODEL, dimension: config.dimension ?? LOCAL_DIMENSION};
}
if (config.kind === 'remote') {
return {kind: 'remote', model: config.model ?? '', dimension: config.dimension ?? 1536};
}
// The union stops literal callers, but `kind` can arrive from untyped
// (JSON-sourced) configuration, and an unknown value must fail here rather than
// falling through to a provider the caller did not ask for.
throw new Error(`embedding: unknown kind '${String((config as {kind: unknown}).kind)}'`);
}
/**
* The shipped loader: a real dynamic import, so the optional package is not required to
* install.
*
* Exported because it is the one thing a capability probe has to exercise directly: calling
* it imports the package — which is cheap, and downloads nothing — while calling `embed`
* would fetch and compile the model. A probe that has to go through `embed` to reach the
* loader either performs network I/O in CI or leaves the loader uncovered, and neither is
* an acceptable answer to "is the optional dependency usable here?".
*/
export async function loadTransformers(): Promise<TransformersModule> {
const moduleName = LOCAL_PACKAGE;
return (await import(moduleName)) as TransformersModule;
}
/**
* ONNX-backed local model (PRD R5), loaded lazily so the package installs and
* runs without the optional dependency or its model present.
*
* The loader is injected rather than imported inline for the same reason the
* remote provider's `fetch` is: the path must be exercisable without the
* heavyweight dependency. A loader that cannot resolve is reported as a named,
* actionable error — the package is optional and undeclared, so "not installed"
* is an expected state rather than an opaque module-resolution crash.
*/
class LocalProviderAdapter implements EmbeddingProvider {
readonly dimension: number;
private readonly load: () => Promise<TransformersModule>;
private extract?: (text: string) => Promise<number[]>;
constructor(dimension: number, load: () => Promise<TransformersModule>) {
this.dimension = dimension;
this.load = load;
}
private async extractor(): Promise<(text: string) => Promise<number[]>> {
if (this.extract) {
return this.extract;
}
let transformers: TransformersModule;
try {
transformers = await this.load();
} catch (cause) {
throw new Error(
`the 'local' embedding provider requires the optional package '${LOCAL_PACKAGE}', ` +
`which could not be loaded (use kind: 'offline', or a remote endpoint): ${String(cause)}`,
{cause},
);
}
const pipeline = await transformers.pipeline('feature-extraction', LOCAL_MODEL);
const extract = async (text: string): Promise<number[]> => {
const output = await pipeline(text, {pooling: 'mean', normalize: true});
return Array.from(output.data);
};
// Stored only once the pipeline is ready, so a failed load is retried
// rather than remembered as a permanent failure.
this.extract = extract;
return extract;
}
async embed(text: string): Promise<number[]> {
return (await this.extractor())(text);
}
}
/**
* The output width of the local model, and the model itself (R5).
*
* Exported because a test that runs the real model has to assert the numbers the
* documentation promises rather than whatever the model happens to return.
*/
export const LOCAL_DIMENSION = 384;
/**
* The embedding a process environment asks for (D24).
*
* The shipped default stays `offline`, in CI and in production: it is deterministic, needs
* no model download and no network, and every test that asserts a *ranking* is written
* against something that cannot change under it. A developer who wants real vectors asks
* for them explicitly — `SEMANTIC_LOG_EMBEDDING=local` — and the same variable is honoured
* by the service and by the flow runner, so one setting makes a local session real.
*
* `remote` is deliberately *not* reachable from here: it needs a url, a model and usually a
* key, and a process that reads three variables to build an endpoint is a configuration file
* wearing an environment variable's clothes.
*/
export function embeddingFromEnv(env: NodeJS.ProcessEnv = process.env): EmbeddingConfig {
const kind = env.SEMANTIC_LOG_EMBEDDING;
if (kind === undefined || kind === '' || kind === 'offline') {
return {kind: 'offline'};
}
if (kind === 'local') {
return {kind: 'local'};
}
throw new Error(`SEMANTIC_LOG_EMBEDDING: unknown kind '${kind}' (use 'offline' or 'local')`);
}
/**
* Build the configured provider. Throws early on an unusable configuration —
* including a `kind` that is not one of the three, and a remote endpoint with no
* url or model. The dimension is taken from the same identity the snapshot records,
* so the vectors and the label they are stored under cannot disagree.
*/
export function createProvider(config: EmbeddingConfig): EmbeddingProvider {
const identity = providerIdentity(config);
if (identity.kind === 'offline') {
return new OfflineProvider(identity.dimension);
}
if (identity.kind === 'remote') {
if (!config.url || !config.model) {
throw new Error('createProvider: remote embedding requires url and model');
}
return new RemoteProvider(
{
url: config.url,
model: config.model,
apiKey: config.apiKey,
fetch: config.fetch ?? globalThis.fetch,
},
identity.dimension,
);
}
return new LocalProviderAdapter(
identity.dimension,
config.loadTransformers ?? loadTransformers,
);
}
|