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
@@ -5,7 +5,7 @@ import { piEngineForSpec, type PiEngine } from "../agent/models.ts";
|
||||
import { createJobRunner } from "../agent/subagents/runner.ts";
|
||||
import { createExtractionPipeline } from "../agent/subagents/extraction.ts";
|
||||
import { runGraphContextJob, type GraphContextDeps } from "../agent/subagents/graph-context.ts";
|
||||
import { EMBEDDING_DIM, EMBEDDING_MODEL, embedNewNodes, embedQuery } from "./embeddings.ts";
|
||||
import { embedNewNodes, embeddingPort } from "./embeddings.ts";
|
||||
import type { ModelRole } from "../../config.ts";
|
||||
import { claimSteers, deliverJobResult, setContextRespawn, setJobRespawn, setTurnRunner } from "./jobs.ts";
|
||||
import { recordToPiMessages } from "./pi/record-adapter.ts";
|
||||
@@ -63,8 +63,8 @@ const jobRunner = createJobRunner({
|
||||
* (/api/graph/import) hit the same single-flight registry. */
|
||||
export const extraction = createExtractionPipeline({
|
||||
...stores,
|
||||
// The REAL embedder (local transformer model): post-merge embedding
|
||||
// batch — semantic clusters are computed and persisted here.
|
||||
// The provider-embedder batch (BYOK wave 6): post-merge embedding —
|
||||
// resolves the model fresh per call; null model = info-skip inside.
|
||||
embedder: embedNewNodes,
|
||||
// BYOK (owner, 2026-09-22): the extraction stages resolve the chain —
|
||||
// role override → slot → NOTHING — read FRESH per stage call; an empty
|
||||
@@ -75,14 +75,27 @@ export const extraction = createExtractionPipeline({
|
||||
});
|
||||
|
||||
/** Graph context injection (vision-context-injection.md): the runner's
|
||||
* deps — same storage ports, the REAL embedder as the query seam (same
|
||||
* warm singleton as extraction, so vector-space parity holds by
|
||||
* construction), and the wake pipeline as the delivery seam. */
|
||||
* deps — same storage ports, the provider-embedder RESOLVER as the query
|
||||
* seam (BYOK wave 6: fresh setting + key reads per run, so vector-space
|
||||
* parity with the ledger is checked against the CURRENT model; a key or
|
||||
* settings PUT between dispatch and spawn lands), and the wake pipeline
|
||||
* as the delivery seam. */
|
||||
export const graphContext: GraphContextDeps = {
|
||||
chatStore: stores.chatStore,
|
||||
jobStore: stores.jobStore,
|
||||
graphStore: stores.graphStore,
|
||||
embedder: { model: EMBEDDING_MODEL, dim: EMBEDDING_DIM, embedQuery },
|
||||
embedder: () => {
|
||||
const resolved = embeddingPort();
|
||||
if (!resolved.ok) return null;
|
||||
const port = resolved.port;
|
||||
return {
|
||||
model: port.model,
|
||||
embedQuery: async (text: string) => {
|
||||
const result = await port.embed([text], "query");
|
||||
return { vector: result.vectors[0]!, promptTokens: result.promptTokens };
|
||||
},
|
||||
};
|
||||
},
|
||||
deliver: deliverJobResult,
|
||||
};
|
||||
|
||||
@@ -221,7 +234,10 @@ export async function handleChatRequest(
|
||||
// GRAPH GATES (settings-storage wave, 2026-09-19): the read channel
|
||||
// needs BOTH the master switch and the embedding rung (injection is a
|
||||
// vector query); with graph off there is no context job at all.
|
||||
if (graphContextEnabled(effective.graph)) {
|
||||
// BYOK (wave 6): an unconfigured embedding model means no vectors to
|
||||
// query — the dispatch skips honestly (the runner's own resolution is
|
||||
// the belt-and-suspenders for respawn races).
|
||||
if (graphContextEnabled(effective.graph) && effective.graph.embeddingModel !== null) {
|
||||
triggerGraphContext(graphContext, chatId);
|
||||
}
|
||||
|
||||
|
||||
+391
-87
@@ -1,94 +1,293 @@
|
||||
/**
|
||||
* THE EMBEDDER (DECIDED 2026-09-13): semantic topic clustering needs
|
||||
* vectors, and vectors need a model running ON the user's machine — no
|
||||
* external services, no API bills, no network dependency for re-embedding.
|
||||
* BYOK PROVIDER EMBEDDINGS (wave 6, 2026-09-23): the local transformers
|
||||
* embedder is dead — vectors now come from the user's keyed provider
|
||||
* (OpenRouter / OpenAI / Google), named by graph.embeddingModel
|
||||
* ("provider:id", null = honestly off; skip with an info log, the graph
|
||||
* stays textual — owner ruling 2026-09-23).
|
||||
*
|
||||
* Runtime: in-process transformers.js (@huggingface/transformers v4 over
|
||||
* onnxruntime-node — native Apple Silicon acceleration). The model
|
||||
* weights download ONCE to the local HF cache on first run (~300MB).
|
||||
* Owner decisions baked in:
|
||||
* - NO factory model (BYOK, like the model slots): null is the floor.
|
||||
* - Adapters resolve FRESH per call from the setting + the auth store
|
||||
* (the search port's treatment — key rotations land immediately); a
|
||||
* key missing for a SET model is an honest error, never a fallback.
|
||||
* - Vectors stored as the provider returns them (default dims, no
|
||||
* truncation, no normalization). The ledger is self-describing on
|
||||
* model + dim; a model change voids it (full re-embed, minutes at
|
||||
* current scale — the local embedder's same staleness rule).
|
||||
* - Usage (prompt tokens) flows to the money channel: every API call
|
||||
* reports one CallUsage step (outputTokens: 0 — embeddings emit no
|
||||
* output; priced via the hand-maintained EMBEDDING_PRICES overlay).
|
||||
*
|
||||
* Model: EmbeddingGemma-300M (q8) — best quality-per-megabyte small
|
||||
* embedding model as of late 2025 (MTEB-en 68.36), built for on-device
|
||||
* use, with MRL truncation so we store 512 of its 768 dims. Every input
|
||||
* uses the model card's document format ("title: none | text: …"),
|
||||
* consistently — mixing prefixed and unprefixed inputs degrades quality.
|
||||
*
|
||||
* Model swaps re-embed EVERYTHING: vectors from different models live in
|
||||
* incompatible spaces. The ledger records model + dim; a mismatch marks
|
||||
* the whole ledger stale and this module re-embeds from scratch (minutes
|
||||
* at current scale).
|
||||
* The adapters are injected-fetch dependencies so tests never touch the
|
||||
* network; wire protocols VERIFIED against provider docs 2026-09-23:
|
||||
* - OpenRouter: POST /api/v1/embeddings (Bearer), input_type
|
||||
* search_document|search_query; catalog GET /api/v1/embeddings/models
|
||||
* (VERIFIED live 2026-09-23 — the inverse path /api/v1/models/
|
||||
* embeddings 404s).
|
||||
* - OpenAI: POST /v1/embeddings (Bearer); catalog GET /v1/models.
|
||||
* - Google: POST /v1beta/models/{m}:batchEmbedContents (x-goog-api-key),
|
||||
* taskType RETRIEVAL_DOCUMENT|RETRIEVAL_QUERY; catalog GET /v1beta/models
|
||||
* filtered on supportedGenerationMethods ∋ embedContent.
|
||||
*/
|
||||
import * as fs from "node:fs";
|
||||
import { env, pipeline, type FeatureExtractionPipeline } from "@huggingface/transformers";
|
||||
import { PATHS } from "../../config.ts";
|
||||
import type { EmbeddingsLedger, GraphNode } from "../core/contracts.ts";
|
||||
|
||||
import type { CallUsage, EmbeddingsLedger, GraphNode } from "../core/contracts.ts";
|
||||
import type { ExtractionStageUsage } from "../core/contracts/chat.ts";
|
||||
import type { EmbeddingPort, EmbeddingResult, EmbeddingTask } from "../core/ports.ts";
|
||||
import { semanticClusters, SEMANTIC_SIM_THRESHOLD } from "../core/semantic-clusters.ts";
|
||||
import { keyValue } from "./keys.ts";
|
||||
import { effectiveSettings } from "./stores.ts";
|
||||
|
||||
export const EMBEDDING_MODEL = "onnx-community/embeddinggemma-300m-ONNX";
|
||||
/** MRL truncation of the native 768 — negligible quality loss, half the
|
||||
* ledger size (2026-09-13 researcher sweep). */
|
||||
export const EMBEDDING_DIM = 512;
|
||||
/** Batch size bounds peak memory; propositions are tiny so 32 per call
|
||||
* keeps the pipeline warm without thrashing. */
|
||||
/** The wire-callable shape the adapters close over (tests inject a stub). */
|
||||
type FetchLike = (input: string, init: RequestInit) => Promise<Response>;
|
||||
|
||||
/** One provider embedding API call — kept small enough to never trip
|
||||
* per-request input limits (propositions are tiny; 32 was the local
|
||||
* embedder's batch and it stays). */
|
||||
const EMBED_BATCH = 32;
|
||||
/** Per-request timeout: a hung provider call must not hang an extraction
|
||||
* run (the wave-1 hung-job lesson). */
|
||||
const REQUEST_TIMEOUT_MS = 60_000;
|
||||
|
||||
// Model cache routing (Electron seam, 2026-09-17): the default cacheDir
|
||||
// lives inside node_modules — read-only inside a packaged .app bundle.
|
||||
// When LH_DATA_DIR is set the cache follows the data seam (userData);
|
||||
// in dev it stays at the package default (the model is already there).
|
||||
if (PATHS.modelCache) {
|
||||
fs.mkdirSync(PATHS.modelCache, { recursive: true });
|
||||
env.cacheDir = PATHS.modelCache;
|
||||
/** The providers this wave ships + their auth vars (the curated key
|
||||
* catalog, same ids the auth store and /api/keys use). */
|
||||
const PROVIDER_KEYS = {
|
||||
openrouter: "OPENROUTER_API_KEY",
|
||||
openai: "OPENAI_API_KEY",
|
||||
google: "GEMINI_API_KEY",
|
||||
} as const;
|
||||
export type EmbeddingProviderId = keyof typeof PROVIDER_KEYS;
|
||||
|
||||
/** "provider:id" → {provider, id}; null when malformed (no guessing). */
|
||||
export function parseEmbeddingModelSlug(slug: string): { provider: EmbeddingProviderId; id: string } | null {
|
||||
const colon = slug.indexOf(":");
|
||||
if (colon <= 0 || colon === slug.length - 1) return null;
|
||||
const provider = slug.slice(0, colon) as EmbeddingProviderId;
|
||||
if (!(provider in PROVIDER_KEYS)) return null;
|
||||
return { provider, id: slug.slice(colon + 1) };
|
||||
}
|
||||
|
||||
let pipelinePromise: Promise<FeatureExtractionPipeline> | null = null;
|
||||
// ─── Adapters (one per provider; pure injected-fetch factories) ──────────────
|
||||
|
||||
function getPipeline(): Promise<FeatureExtractionPipeline> {
|
||||
pipelinePromise ??= pipeline("feature-extraction", EMBEDDING_MODEL, { dtype: "q8" });
|
||||
return pipelinePromise;
|
||||
/** Shared: HTTP-error → clear Error (the search adapters' mapping). */
|
||||
async function demandOk(resp: Response, label: string): Promise<void> {
|
||||
if (resp.ok) return;
|
||||
const body = await resp.text();
|
||||
throw new Error(`${label} API ${resp.status}: ${body.slice(0, 200)}`);
|
||||
}
|
||||
|
||||
/** EmbeddingGemma's required document format (model card). */
|
||||
function embedInput(statement: string): string {
|
||||
return `title: none | text: ${statement}`;
|
||||
/** Shared: normalize a provider's {embedding, index} rows to input order. */
|
||||
function rowsInOrder(
|
||||
data: Array<{ embedding?: unknown; index?: unknown }>,
|
||||
count: number,
|
||||
label: string,
|
||||
): number[][] {
|
||||
const byIndex = new Map<number, number[]>();
|
||||
for (const row of data) {
|
||||
if (
|
||||
Array.isArray(row.embedding) &&
|
||||
row.embedding.every((x) => typeof x === "number") &&
|
||||
typeof row.index === "number"
|
||||
) {
|
||||
byIndex.set(row.index, row.embedding as number[]);
|
||||
}
|
||||
}
|
||||
const out: number[][] = [];
|
||||
for (let i = 0; i < count; i++) {
|
||||
const vec = byIndex.get(i);
|
||||
if (!vec) throw new Error(`${label} API returned ${byIndex.size}/${count} vectors`);
|
||||
out.push(vec);
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
/** EmbeddingGemma's retrieval-QUERY format (model card) — the topic side of
|
||||
* graph context injection (vision-context-injection.md). Role-asymmetric
|
||||
* BY DESIGN: queries and documents use different trained templates; the
|
||||
* requirement is same model + same truncation (one vector space), each
|
||||
* side in the format the model was trained for. */
|
||||
function queryInput(text: string): string {
|
||||
return `task: search result | query: ${text}`;
|
||||
/** One API call's money-channel step: embeddings emit no output, cache
|
||||
* nothing, reason never — the fields exist (CallUsage requires them) as
|
||||
* absent-is-not-zero undefineds. */
|
||||
function embeddingStep(promptTokens: number): CallUsage {
|
||||
return {
|
||||
inputTokens: promptTokens,
|
||||
outputTokens: 0,
|
||||
cacheReadTokens: undefined,
|
||||
cacheWriteTokens: undefined,
|
||||
reasoningTokens: undefined,
|
||||
};
|
||||
}
|
||||
|
||||
/** Embed one retrieval query (the session's topic) — same warm singleton,
|
||||
* same 512-dim MRL truncation as the ledger. The one embedding call the
|
||||
* injection job makes; tens–~200ms warm. */
|
||||
export async function embedQuery(text: string): Promise<number[]> {
|
||||
const extractor = await getPipeline();
|
||||
const output = await extractor([queryInput(text)], { pooling: "mean", normalize: true });
|
||||
return truncateNormalize((output.tolist() as number[][])[0]!);
|
||||
/** OpenRouter: POST https://openrouter.ai/api/v1/embeddings — Bearer auth,
|
||||
* input_type hints the retrieval side (verified against the docs 2026-09-23). */
|
||||
export function openrouterEmbeddingProvider(model: string, apiKey: string, fetchFn: FetchLike = fetch): EmbeddingPort {
|
||||
const parsed = parseEmbeddingModelSlug(model);
|
||||
if (!parsed) throw new Error(`invalid embedding model slug: ${model}`);
|
||||
return {
|
||||
model,
|
||||
async embed(texts: readonly string[], task: EmbeddingTask): Promise<EmbeddingResult> {
|
||||
const resp = await fetchFn("https://openrouter.ai/api/v1/embeddings", {
|
||||
method: "POST",
|
||||
headers: { "content-type": "application/json", authorization: `Bearer ${apiKey}` },
|
||||
body: JSON.stringify({
|
||||
model: parsed.id,
|
||||
input: texts,
|
||||
input_type: task === "query" ? "search_query" : "search_document",
|
||||
}),
|
||||
signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),
|
||||
});
|
||||
await demandOk(resp, "OpenRouter embeddings");
|
||||
const data = (await resp.json()) as {
|
||||
data?: Array<{ embedding?: unknown; index?: unknown }>;
|
||||
usage?: { prompt_tokens?: number };
|
||||
};
|
||||
return {
|
||||
vectors: rowsInOrder(data.data ?? [], texts.length, "OpenRouter embeddings"),
|
||||
model,
|
||||
dim: 0, // filled by the caller from the first vector (below)
|
||||
promptTokens: data.usage?.prompt_tokens,
|
||||
};
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
/** Truncate an L2-normalized vector to EMBEDDING_DIM (MRL: leading dims
|
||||
* carry most of the signal) and re-normalize so cosine = dot product. */
|
||||
function truncateNormalize(vec: readonly number[]): number[] {
|
||||
const out = vec.slice(0, EMBEDDING_DIM);
|
||||
let mag = 0;
|
||||
for (const x of out) mag += x * x;
|
||||
mag = Math.sqrt(mag);
|
||||
return mag > 0 ? out.map((x) => x / mag) : out;
|
||||
/** OpenAI: POST https://api.openai.com/v1/embeddings — Bearer auth; the
|
||||
* response's usage carries prompt_tokens (verified 2026-09-23). */
|
||||
export function openaiEmbeddingProvider(model: string, apiKey: string, fetchFn: FetchLike = fetch): EmbeddingPort {
|
||||
const parsed = parseEmbeddingModelSlug(model);
|
||||
if (!parsed) throw new Error(`invalid embedding model slug: ${model}`);
|
||||
return {
|
||||
model,
|
||||
async embed(texts: readonly string[], task: EmbeddingTask): Promise<EmbeddingResult> {
|
||||
// OpenAI's embeddings API takes no task hint (one format per
|
||||
// model); the argument exists for seam parity and is unused here.
|
||||
void task;
|
||||
const resp = await fetchFn("https://api.openai.com/v1/embeddings", {
|
||||
method: "POST",
|
||||
headers: { "content-type": "application/json", authorization: `Bearer ${apiKey}` },
|
||||
body: JSON.stringify({ model: parsed.id, input: texts }),
|
||||
signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),
|
||||
});
|
||||
await demandOk(resp, "OpenAI embeddings");
|
||||
const data = (await resp.json()) as {
|
||||
data?: Array<{ embedding?: unknown; index?: unknown }>;
|
||||
usage?: { prompt_tokens?: number };
|
||||
};
|
||||
return {
|
||||
vectors: rowsInOrder(data.data ?? [], texts.length, "OpenAI embeddings"),
|
||||
model,
|
||||
dim: 0,
|
||||
promptTokens: data.usage?.prompt_tokens,
|
||||
};
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
/** Google: POST …/models/{m}:batchEmbedContents — x-goog-api-key, one
|
||||
* request per text with taskType RETRIEVAL_QUERY|RETRIEVAL_DOCUMENT
|
||||
* (verified against the Gemini API docs 2026-09-23). */
|
||||
export function googleEmbeddingProvider(model: string, apiKey: string, fetchFn: FetchLike = fetch): EmbeddingPort {
|
||||
const parsed = parseEmbeddingModelSlug(model);
|
||||
if (!parsed) throw new Error(`invalid embedding model slug: ${model}`);
|
||||
return {
|
||||
model,
|
||||
async embed(texts: readonly string[], task: EmbeddingTask): Promise<EmbeddingResult> {
|
||||
const resp = await fetchFn(
|
||||
`https://generativelanguage.googleapis.com/v1beta/models/${encodeURIComponent(parsed.id)}:batchEmbedContents`,
|
||||
{
|
||||
method: "POST",
|
||||
headers: { "content-type": "application/json", "x-goog-api-key": apiKey },
|
||||
body: JSON.stringify({
|
||||
requests: texts.map((text) => ({
|
||||
model: `models/${parsed.id}`,
|
||||
content: { parts: [{ text }] },
|
||||
taskType: task === "query" ? "RETRIEVAL_QUERY" : "RETRIEVAL_DOCUMENT",
|
||||
})),
|
||||
}),
|
||||
signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),
|
||||
},
|
||||
);
|
||||
await demandOk(resp, "Google embeddings");
|
||||
const data = (await resp.json()) as {
|
||||
embeddings?: Array<{ values?: unknown }>;
|
||||
/** embedContent responses carry usageMetadata; the batch
|
||||
* shape may too — parsed opportunistically, never guessed. */
|
||||
usageMetadata?: { promptTokenCount?: number };
|
||||
};
|
||||
const vectors = (data.embeddings ?? []).map((row, i) => {
|
||||
if (!Array.isArray(row.values) || !row.values.every((x) => typeof x === "number")) {
|
||||
throw new Error(`Google embeddings API returned a bad vector at index ${i}`);
|
||||
}
|
||||
return row.values as number[];
|
||||
});
|
||||
if (vectors.length !== texts.length) {
|
||||
throw new Error(`Google embeddings API returned ${vectors.length}/${texts.length} vectors`);
|
||||
}
|
||||
return { vectors, model, dim: 0, promptTokens: data.usageMetadata?.promptTokenCount };
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
// ─── Resolution (fresh per call — the search port's treatment) ───────────────
|
||||
|
||||
/** The resolved state: a live port, or the honest reason it isn't. */
|
||||
export type EmbeddingResolution =
|
||||
| { ok: true; port: EmbeddingPort }
|
||||
| { ok: false; reason: string };
|
||||
|
||||
/** Pure resolution from a setting slug + a key lookup (testable): null
|
||||
* slug = off; a set slug with no key = honest unavailability. */
|
||||
export function resolveEmbeddingPortFor(
|
||||
slug: string | null | undefined,
|
||||
keyOf: (envVar: string) => string | undefined,
|
||||
fetchFn?: FetchLike,
|
||||
): EmbeddingResolution {
|
||||
if (slug === null || slug === undefined || slug === "") {
|
||||
return { ok: false, reason: "no embedding model configured (graph.embeddingModel is null)" };
|
||||
}
|
||||
const parsed = parseEmbeddingModelSlug(slug);
|
||||
if (!parsed) {
|
||||
return { ok: false, reason: `invalid embedding model slug: ${slug}` };
|
||||
}
|
||||
const apiKey = keyOf(PROVIDER_KEYS[parsed.provider]);
|
||||
if (!apiKey) {
|
||||
return {
|
||||
ok: false,
|
||||
reason: `embedding model is set to "${slug}" but its ${PROVIDER_KEYS[parsed.provider]} key is missing`,
|
||||
};
|
||||
}
|
||||
const factory =
|
||||
parsed.provider === "openrouter"
|
||||
? openrouterEmbeddingProvider
|
||||
: parsed.provider === "openai"
|
||||
? openaiEmbeddingProvider
|
||||
: googleEmbeddingProvider;
|
||||
return { ok: true, port: factory(slug, apiKey, fetchFn) };
|
||||
}
|
||||
|
||||
/** The LIVE resolution — reads the settings record + auth store fresh on
|
||||
* every call, so a settings PUT or a key save lands on the next call. */
|
||||
export function embeddingPort(): EmbeddingResolution {
|
||||
return resolveEmbeddingPortFor(effectiveSettings().graph.embeddingModel, keyValue);
|
||||
}
|
||||
|
||||
// ─── The post-merge batch (the extraction pipeline's embedding tail) ────────
|
||||
|
||||
/** What the embedding tail reports back: counts for logs, the skip reason
|
||||
* when nothing ran (owner ruling: null model = info skip, not error), and
|
||||
* the money-channel stage when it billed (one CallUsage step per API call,
|
||||
* outputTokens: 0 — priced via EMBEDDING_PRICES). */
|
||||
export interface EmbedBatchResult {
|
||||
embedded: number;
|
||||
total: number;
|
||||
skipped?: string;
|
||||
usage?: ExtractionStageUsage;
|
||||
}
|
||||
|
||||
/**
|
||||
* Embed every node the ledger is missing — the post-merge hook (new
|
||||
* propositions only, typically a handful) AND the cold-start backfill
|
||||
* (empty ledger = every node) in one function. Model/dim mismatch voids
|
||||
* the ledger first (full re-embed). Also prunes vectors for nodes that
|
||||
* no longer exist. Writes the whole ledger and returns counts for logs.
|
||||
* Non-fatal by contract: the caller treats failure as "clustering runs
|
||||
* without the new nodes until the next merge retries".
|
||||
* (empty ledger = every node) in one function. A model change voids the
|
||||
* ledger first (full re-embed — vectors from different models live in
|
||||
* incompatible spaces). Also prunes vectors for nodes that no longer
|
||||
* exist, then re-clusters ONCE per batch (consumers read the persisted
|
||||
* partition). Non-fatal by contract: the caller treats failure as
|
||||
* "clustering runs without the new nodes until the next merge retries".
|
||||
*/
|
||||
export async function embedNewNodes(
|
||||
graphStore: {
|
||||
@@ -96,10 +295,21 @@ export async function embedNewNodes(
|
||||
readEmbeddings(): EmbeddingsLedger;
|
||||
writeEmbeddings(ledger: EmbeddingsLedger): void;
|
||||
},
|
||||
): Promise<{ embedded: number; total: number }> {
|
||||
/** Test seam: the live resolution by default; tests inject a scripted
|
||||
* port (they never read the machine's settings/auth files). */
|
||||
resolvePort: () => EmbeddingResolution = embeddingPort,
|
||||
): Promise<EmbedBatchResult> {
|
||||
const resolved = resolvePort();
|
||||
const ledger = graphStore.readEmbeddings();
|
||||
const stale = ledger.model !== EMBEDDING_MODEL || ledger.dim !== EMBEDDING_DIM;
|
||||
if (!resolved.ok) {
|
||||
console.info(`[embeddings] skipped — ${resolved.reason}`);
|
||||
return { embedded: 0, total: Object.keys(ledger.vectors).length, skipped: resolved.reason };
|
||||
}
|
||||
const port = resolved.port;
|
||||
const stale = ledger.model !== port.model;
|
||||
const vectors: Record<string, number[]> = stale ? {} : { ...ledger.vectors };
|
||||
const steps: NonNullable<EmbedBatchResult["usage"]>["steps"] = [];
|
||||
let dim = ledger.dim;
|
||||
|
||||
// Node LINES may repeat an id — dedupe, first defines (getGraph rule).
|
||||
const seen = new Set<string>();
|
||||
@@ -110,18 +320,20 @@ export async function embedNewNodes(
|
||||
if (!vectors[node.id]) todo.push(node);
|
||||
}
|
||||
|
||||
if (todo.length > 0) {
|
||||
const extractor = await getPipeline();
|
||||
for (let i = 0; i < todo.length; i += EMBED_BATCH) {
|
||||
const batch = todo.slice(i, i + EMBED_BATCH);
|
||||
const output = await extractor(batch.map((n) => embedInput(n.statement)), {
|
||||
pooling: "mean",
|
||||
normalize: true,
|
||||
});
|
||||
const rows = output.tolist() as number[][];
|
||||
for (let j = 0; j < batch.length; j++) {
|
||||
vectors[batch[j]!.id] = truncateNormalize(rows[j]!);
|
||||
}
|
||||
for (let i = 0; i < todo.length; i += EMBED_BATCH) {
|
||||
const batch = todo.slice(i, i + EMBED_BATCH);
|
||||
const result = await port.embed(
|
||||
batch.map((n) => n.statement),
|
||||
"document",
|
||||
);
|
||||
for (let j = 0; j < batch.length; j++) {
|
||||
vectors[batch[j]!.id] = result.vectors[j]!;
|
||||
}
|
||||
if (result.vectors[0]) dim = result.vectors[0].length;
|
||||
if (result.promptTokens !== undefined) {
|
||||
// outputTokens: 0 — embeddings emit no output; callCost needs it
|
||||
// defined to price the input at all.
|
||||
steps.push(embeddingStep(result.promptTokens));
|
||||
}
|
||||
}
|
||||
|
||||
@@ -135,20 +347,112 @@ export async function embedNewNodes(
|
||||
}
|
||||
|
||||
// RECLUSTER AT THE END OF THE BATCH (user-decided 2026-09-13): the
|
||||
// agglomerative pass is ~1.5s at current scale, so it runs HERE —
|
||||
// once per embedding batch — and consumers (snapshot, settle, namer)
|
||||
// read the persisted partition instead of recomputing per read.
|
||||
// agglomerative pass runs HERE — once per embedding batch — and
|
||||
// consumers (snapshot, settle, namer) read the persisted partition.
|
||||
const clusters = semanticClusters(vectors, { similarityThreshold: SEMANTIC_SIM_THRESHOLD });
|
||||
|
||||
const wrote = todo.length > 0 || pruned > 0 || stale || ledger.clusters.length !== clusters.length;
|
||||
if (wrote) {
|
||||
graphStore.writeEmbeddings({
|
||||
model: EMBEDDING_MODEL,
|
||||
dim: EMBEDDING_DIM,
|
||||
model: port.model,
|
||||
dim,
|
||||
vectors,
|
||||
clusters,
|
||||
updatedAt: new Date().toISOString(),
|
||||
});
|
||||
}
|
||||
return { embedded: todo.length, total: Object.keys(vectors).length };
|
||||
return {
|
||||
embedded: todo.length,
|
||||
total: Object.keys(vectors).length,
|
||||
usage: steps.length > 0 ? { model: port.model, steps } : undefined,
|
||||
};
|
||||
}
|
||||
// ─── The federated embedding catalog (wave 6, owner ruling 2026-09-23) ──────
|
||||
|
||||
/** One catalog row: provider-qualified id (the setting's exact slug
|
||||
* format) + display name. */
|
||||
export interface EmbeddingCatalogModel {
|
||||
/** The full "provider:id" slug — what graph.embeddingModel stores. */
|
||||
id: string;
|
||||
/** Human label for the picker (the provider's name, when given). */
|
||||
name: string;
|
||||
}
|
||||
|
||||
/** The embedding models of every KEYED provider, fetched live (the
|
||||
* chat-models picker is a static pi-ai catalog, but pi-ai carries no
|
||||
* embedding models — so this endpoint federates the providers' own
|
||||
* lists). A provider without a key contributes nothing; a provider whose
|
||||
* catalog call fails is skipped with an error log (never poisons the
|
||||
* rest, never fails the endpoint). Empty array = nothing keyed. */
|
||||
export async function embeddingModelCatalog(
|
||||
fetchFn: FetchLike = fetch,
|
||||
/** Test seam: the live key lookup by default; tests inject scripted
|
||||
* keys (they never read the machine's auth file). */
|
||||
keyOf: (id: string) => string | undefined = keyValue,
|
||||
): Promise<EmbeddingCatalogModel[]> {
|
||||
const out: EmbeddingCatalogModel[] = [];
|
||||
const fetchJson = async (url: string, headers: Record<string, string>): Promise<unknown> => {
|
||||
const resp = await fetchFn(url, { headers, signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS) });
|
||||
await demandOk(resp, "Embeddings catalog");
|
||||
return resp.json();
|
||||
};
|
||||
|
||||
// OpenRouter: GET /api/v1/embeddings/models — the dedicated embedding
|
||||
// list (VERIFIED live 2026-09-23: the inverse path /api/v1/models/
|
||||
// embeddings 404s; this one returns {data: [{id, name, …}]}). Paginated
|
||||
// — limit=500 covers the whole list at current size (33).
|
||||
if (keyOf(PROVIDER_KEYS.openrouter)) {
|
||||
try {
|
||||
const data = (await fetchJson("https://openrouter.ai/api/v1/embeddings/models?limit=500", {
|
||||
authorization: `Bearer ${keyOf(PROVIDER_KEYS.openrouter)}`,
|
||||
})) as { data?: Array<{ id?: unknown; name?: unknown }> };
|
||||
for (const m of data.data ?? []) {
|
||||
if (typeof m.id !== "string" || !m.id) continue;
|
||||
out.push({ id: `openrouter:${m.id}`, name: typeof m.name === "string" && m.name ? m.name : m.id });
|
||||
}
|
||||
} catch (err) {
|
||||
console.error("[embeddings] OpenRouter catalog fetch failed:", err);
|
||||
}
|
||||
}
|
||||
|
||||
// OpenAI: GET /v1/models — no embedding-only list, so filter the
|
||||
// generic one on the embedding naming convention.
|
||||
if (keyOf(PROVIDER_KEYS.openai)) {
|
||||
try {
|
||||
const data = (await fetchJson("https://api.openai.com/v1/models", {
|
||||
authorization: `Bearer ${keyOf(PROVIDER_KEYS.openai)}`,
|
||||
})) as { data?: Array<{ id?: unknown }> };
|
||||
for (const m of data.data ?? []) {
|
||||
if (typeof m.id !== "string" || !/^text-embedding-/.test(m.id)) continue;
|
||||
out.push({ id: `openai:${m.id}`, name: m.id });
|
||||
}
|
||||
} catch (err) {
|
||||
console.error("[embeddings] OpenAI catalog fetch failed:", err);
|
||||
}
|
||||
}
|
||||
|
||||
// Google: GET /v1beta/models — filter on the methods the model
|
||||
// actually serves (embedContent present).
|
||||
if (keyOf(PROVIDER_KEYS.google)) {
|
||||
try {
|
||||
const data = (await fetchJson("https://generativelanguage.googleapis.com/v1beta/models?pageSize=1000", {
|
||||
"x-goog-api-key": keyOf(PROVIDER_KEYS.google)!,
|
||||
})) as {
|
||||
models?: Array<{ name?: unknown; displayName?: unknown; supportedGenerationMethods?: unknown }>;
|
||||
};
|
||||
for (const m of data.models ?? []) {
|
||||
if (typeof m.name !== "string" || !m.name.startsWith("models/")) continue;
|
||||
if (!Array.isArray(m.supportedGenerationMethods) || !m.supportedGenerationMethods.includes("embedContent")) continue;
|
||||
const id = m.name.slice("models/".length);
|
||||
out.push({
|
||||
id: `google:${id}`,
|
||||
name: typeof m.displayName === "string" && m.displayName ? m.displayName : id,
|
||||
});
|
||||
}
|
||||
} catch (err) {
|
||||
console.error("[embeddings] Google catalog fetch failed:", err);
|
||||
}
|
||||
}
|
||||
|
||||
return out;
|
||||
}
|
||||
|
||||
@@ -6,6 +6,7 @@
|
||||
|
||||
export * from "./chat-handler.ts";
|
||||
export * from "./chats.ts";
|
||||
export * from "./embeddings.ts";
|
||||
export * from "./graph.ts";
|
||||
export * from "./commit-positions.ts";
|
||||
export * from "./jobs.ts";
|
||||
|
||||
@@ -277,7 +277,11 @@ export function parseAppearanceSettingsBody(body: unknown): SettingsBodyResult<A
|
||||
return { ok: true, settings };
|
||||
}
|
||||
|
||||
/** graph: the four-rung ladder { enabled?, embedding?, clustering?, labeling? }. */
|
||||
/** graph: the four-rung ladder { enabled?, embedding?, clustering?, labeling? }
|
||||
* + embeddingModel? — the BYOK provider embedding model (wave 6,
|
||||
* 2026-09-23): null or a "provider:id" slug of a shipped provider.
|
||||
* Catalog validity is network-dependent — the parser checks SHAPE only
|
||||
* (the federated /api/models/embeddings gives the picker real options). */
|
||||
export function parseGraphSettingsBody(body: unknown): SettingsBodyResult<GraphSettings> {
|
||||
const raw = asRecord(body);
|
||||
if (!raw) return { ok: false, error: "Invalid JSON body" };
|
||||
@@ -290,6 +294,25 @@ export function parseGraphSettingsBody(body: unknown): SettingsBodyResult<GraphS
|
||||
settings[key] = raw[key];
|
||||
}
|
||||
}
|
||||
if (raw.embeddingModel !== undefined) {
|
||||
if (raw.embeddingModel !== null) {
|
||||
// NOTE: the id may itself contain colons (OpenRouter suffixes like
|
||||
// "liquid/lfm-2.5-embedding-350m:free") — the provider is the part
|
||||
// before the FIRST colon, so the id half just forbids whitespace.
|
||||
if (
|
||||
typeof raw.embeddingModel !== "string" ||
|
||||
!/^(openrouter|openai|google):\S+$/.test(raw.embeddingModel)
|
||||
) {
|
||||
return {
|
||||
ok: false,
|
||||
error: 'embeddingModel must be null or a "openrouter|openai|google:model-id" slug',
|
||||
};
|
||||
}
|
||||
settings.embeddingModel = raw.embeddingModel;
|
||||
} else {
|
||||
settings.embeddingModel = null;
|
||||
}
|
||||
}
|
||||
return { ok: true, settings };
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user