BYOK provider embeddings (wave 6 chunk B): kill local transformers

Replace the @huggingface/transformers local embedder (~316MB ONNX cache)
with an EmbeddingPort seam: graph.embeddingModel (nullable "provider:id"
slug) selects an OpenRouter/OpenAI/Gemini adapter, resolved fresh per
call from settings + the auth store so key rotations land immediately.

- src/server/embeddings.ts rewritten: slug grammar, injected-fetch
  adapters (OpenRouter input_type, Gemini taskType batching, OpenAI
  index-ordered rows), honest resolution (null = skip-with-info, set
  model with missing key = honest error), embedNewNodes with model-change
  re-embed wipe + ghost pruning + usage steps, federated catalog
  (OpenRouter /api/v1/embeddings/models — VERIFIED live, the inverse
  path 404s; OpenAI text-embedding- filter; Gemini embedContent filter).
- GET /api/models/embeddings: federated catalog, empty when no keys.
- Extraction tail records stages.embedding usage into the run's money
  channel; EMBEDDING_PRICES overlay in core/prices.ts (outside the
  sync-generated block), callCost falls back to it; cost-summary folds
  the fourth stage; graph-context injection resolves its embedder per
  run, fails honestly when unconfigured, records query tokens.
- Settings: GraphSettings.embeddingModel (null factory), section parser
  validates slug shape (id may contain colons — OpenRouter :free
  suffixes, live catalog fact).
- Removal: @huggingface/transformers uninstalled, PATHS.modelCache and
  onnx tracing entries deleted, @huggingface tracing-excludes guard
  added, stale sharp/libvips comments updated to historical.

Live-verified on port 3004: catalog federates 36 rows (33 OpenRouter +
3 OpenAI), embedNewNodes embedded 10 nodes via the free liquid/lfm-2.5
model (dim 1024, 435 prompt tokens, 1 call, $0), settings PUT rejects
bad slugs with 400.

Gates: typecheck, lint, 895/895 tests, build.
This commit is contained in:
amosblomqvist
2026-09-23 03:53:57 +03:00
parent a6338511d3
commit 0a4a9a11de
29 changed files with 1233 additions and 841 deletions
+4
View File
@@ -108,6 +108,10 @@ export interface ExtractionRunUsage {
stages: {
extractor: ExtractionStageUsage;
auditor: ExtractionStageUsage;
/** Wave 6 (2026-09-23): the provider-embedding tail — present only
* when an embedding model is configured and the run embedded
* (local embeddings were free, so this post-dates them). */
embedding?: ExtractionStageUsage;
};
}
+25 -7
View File
@@ -55,10 +55,12 @@ export interface GraphCostSummary {
totalDollars: number;
/** Runs where EVERY stage was unpriced (totalDollars excludes them). */
unpricedRuns: number;
/** The three billable pipeline stages (merger/embedder are free). */
/** The billable pipeline stages (wave 6: the embedder became one —
* provider embeddings cost money; merger is the only free stage). */
stages: {
extraction: StageCost;
audit: StageCost;
embedding: StageCost;
naming: StageCost;
};
/** Newest first, capped at `recentCap` (default 10). */
@@ -113,6 +115,7 @@ function mergeTokens(into: CallUsage, steps: readonly CallUsage[]): void {
export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
const extraction = emptyStage();
const audit = emptyStage();
const embedding = emptyStage();
const naming = emptyStage();
const runs: RecentRun[] = [];
@@ -131,10 +134,25 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
if (audDollars === null) audit.unpricedRuns++;
else audit.dollars += audDollars;
// Run-level: null only when BOTH stages are unpriced.
const priced = extDollars ?? 0;
const runDollars = extDollars === null && audDollars === null ? null : priced + (audDollars ?? 0);
const allSteps = [...stages.extractor.steps, ...stages.auditor.steps];
// Wave 6: the embedding tail — present only when an embedding model
// ran (absent stage = the run embedded nothing, never "free").
let embDollars: number | null = null;
let embSteps: CallUsage[] = [];
if (stages.embedding) {
embedding.runs++;
embSteps = stages.embedding.steps;
mergeTokens(embedding.tokens, embSteps);
embDollars = stageDollars(stages.embedding.model, embSteps);
if (embDollars === null) embedding.unpricedRuns++;
else embedding.dollars += embDollars;
}
// Run-level: null only when EVERY stage is unpriced.
const allStages: Array<number | null> = [extDollars, audDollars, embDollars];
const runDollars = allStages.every((d) => d === null)
? null
: allStages.reduce<number>((a, d) => a + (d ?? 0), 0);
const allSteps = [...stages.extractor.steps, ...stages.auditor.steps, ...embSteps];
runs.push({
startedAt: run.startedAt,
at: run.at,
@@ -177,9 +195,9 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
// A run is "unpriced" when it billed nothing countable at all.
const unpricedRuns = runs.filter((r) => r.dollars === null).length;
return {
totalDollars: extraction.dollars + audit.dollars + naming.dollars,
totalDollars: extraction.dollars + audit.dollars + embedding.dollars + naming.dollars,
unpricedRuns,
stages: { extraction, audit, naming },
stages: { extraction, audit, embedding, naming },
runs: capped,
};
}
+46
View File
@@ -243,3 +243,49 @@ export interface AuthStore {
/** Remove one VAR if present (absent = no-op, still a silent success). */
deleteAuthVar(name: string): void;
}
// ─── EmbeddingPort — BYOK provider embeddings (wave 6, 2026-09-23) ───────────
/**
* The provider-embedding seam (the local transformers embedder died with
* this wave): vectors come from the user's keyed provider — OpenRouter,
* OpenAI, or Google — named by graph.embeddingModel ("provider:id", null =
* honestly off: embedding/clustering/injection skip with an info log).
*
* Owner decisions (2026-09-23) baked in:
* - The adapters are resolved FRESH per call from the setting + the auth
* store (key rotations land immediately — the search port's treatment).
* - Vectors are stored as the provider returns them (default dims, no
* truncation, no normalization); the ledger is self-describing on
* model + dim, and any model change voids it (full re-embed).
* - Usage (prompt tokens) flows back to the money channel: adapters
* report it, the caller records it as a CallUsage step.
*
* Pure types only (browser-safe): the adapters are server-side
* (src/server/embeddings.ts) and arrive as injected dependencies.
*/
/** Which side of retrieval the text is — providers bill identically but
* quality-optimize differently (OpenRouter's input_type, Gemini's
* taskType); every input in one call shares the task. */
export type EmbeddingTask = "document" | "query";
/** One embedding batch's result: vectors in input order, the model's
* identity + dim (the ledger stamps), and the provider-reported prompt
* token count (undefined = provider didn't report; cost stays honest
* "—" via the existing unpriced path). */
export interface EmbeddingResult {
vectors: number[][];
/** The full "provider:id" slug — ledger/cost keys use it verbatim. */
model: string;
dim: number;
promptTokens?: number;
}
/** A live provider embedding adapter. `embed` throws on failure (adapters
* map HTTP errors to clear Error messages); callers treat failure as
* non-fatal per the extraction contract. */
export interface EmbeddingPort {
model: string;
embed(texts: readonly string[], task: EmbeddingTask): Promise<EmbeddingResult>;
}
+26 -1
View File
@@ -2752,6 +2752,30 @@ export const CONTEXT_LIMITS: Record<string, number> = {
// ─── Pure helpers (folds + cost inference) ───────────────────────────────────
// ─── Embedding prices (wave 6, 2026-09-23) ─────────────────────────────────
/** Hand-maintained BYOK embedding prices, USD per 1M INPUT tokens — the
* sync script never touches this block (pi-ai's catalogs carry no
* embedding models, so the generated PRICES table has none of these).
* VERIFIED against provider pricing pages 2026-09-23: OpenAI's model
* docs ($0.02 3-small / $0.13 3-large / $0.10 ada-002), Google's pricing
* page ($0.15 gemini-embedding-001 / $0.20 Gemini Embedding 2 text),
* OpenRouter passes provider pricing through. output: 0 — embeddings
* emit no output tokens. Unlisted models stay unpriced ("—", never
* zero); tokens still get captured. */
export const EMBEDDING_PRICES: Record<string, ModelPrice> = {
"openai:text-embedding-3-small": { input: 0.02, output: 0 },
"openai:text-embedding-3-large": { input: 0.13, output: 0 },
"openai:text-embedding-ada-002": { input: 0.1, output: 0 },
"google:gemini-embedding-001": { input: 0.15, output: 0 },
"google:gemini-embedding-2": { input: 0.2, output: 0 },
"openrouter:openai/text-embedding-3-small": { input: 0.02, output: 0 },
"openrouter:openai/text-embedding-3-large": { input: 0.13, output: 0 },
"openrouter:openai/text-embedding-ada-002": { input: 0.1, output: 0 },
"openrouter:google/gemini-embedding-001": { input: 0.15, output: 0 },
};
/** Sum a list of per-step usages into one CallUsage. A field stays
* undefined when EVERY step left it unreported — absent is not zero. */
export function sumSteps(steps: readonly CallUsage[]): CallUsage {
@@ -2773,7 +2797,8 @@ export function callCost(
usage: { inputTokens?: number; outputTokens?: number; cacheReadTokens?: number; cacheWriteTokens?: number },
model: string,
): number | null {
const price = PRICES[model];
// Wave 6: the hand-maintained embedding overlay backs the lookup too.
const price = PRICES[model] ?? EMBEDDING_PRICES[model];
if (!price) return null;
if (usage.inputTokens === undefined || usage.outputTokens === undefined) return null;
const cacheRead = usage.cacheReadTokens ?? 0;
+3
View File
@@ -201,6 +201,9 @@ export function resolveSettings(record: UserSettings | null): EffectiveSettings
embedding: graph.embedding ?? FACTORY_SETTINGS.graph.embedding,
clustering: graph.clustering ?? FACTORY_SETTINGS.graph.clustering,
labeling: graph.labeling ?? FACTORY_SETTINGS.graph.labeling,
// BYOK (wave 6): null is the factory floor and a VALID resolved
// value — ?? only fills the absent case, null passes through.
embeddingModel: graph.embeddingModel ?? FACTORY_SETTINGS.graph.embeddingModel,
},
};
}
+16 -1
View File
@@ -251,6 +251,13 @@ export interface GraphSettings {
clustering?: boolean;
/** The LLM cluster-namer (needs clustering). */
labeling?: boolean;
/** BYOK embeddings (wave 6, 2026-09-23): the provider embedding model
* as a "provider:id" slug (openrouter:…, openai:…, google:…). null =
* no model — embedding/clustering/injection skip honestly (info log,
* graph stays textual); no factory floor exists. Keys live in the auth
* store, resolved fresh per call. A model CHANGE re-embeds everything
* (the ledger's staleness semantics). */
embeddingModel?: string | null;
}
/** THE record. Every section optional (delta); unknown top-level keys are
@@ -303,7 +310,7 @@ export const FACTORY_SETTINGS: FactorySettings = {
sessionNameSerif: true,
mdRendering: { headingWeight: 900, boldColor: null, italicColor: null },
},
graph: { enabled: true, embedding: true, clustering: true, labeling: true },
graph: { enabled: true, embedding: true, clustering: true, labeling: true, embeddingModel: null },
};
// ─── Sanitizer ─────────────────────────────────────────────────────────────
@@ -453,6 +460,14 @@ function sanitizeGraph(raw: Record<string, unknown>): GraphSettings {
if (clustering !== undefined) out.clustering = clustering;
const labeling = boolOrDrop(raw.labeling);
if (labeling !== undefined) out.labeling = labeling;
// BYOK embeddings (wave 6): null or a provider-qualified slug; anything
// else (bad type, bad shape) DROPS to absent — an invalid write never
// becomes a live model (the PUT parser 400s it first; this is the
// read-side belt-and-suspenders for hand-edited files).
if (raw.embeddingModel === null) out.embeddingModel = null;
else if (typeof raw.embeddingModel === "string" && /^\S+:\S+$/.test(raw.embeddingModel)) {
out.embeddingModel = raw.embeddingModel;
}
return out;
}