@voltro/ai 0.28.0 → 0.30.0

package/dist/inferenceQueue-B25RTJZ4.js

@@ -0,0 +1,447 @@
+import { a as e, r as t } from "./provider-BOC5wBxx.js";
+import { Output as n, generateObject as r, generateText as i, jsonSchema as a, stepCountIs as o, tool as s } from "ai";
+import { Context as c, Effect as l, JSONSchema as u, Metric as d, MetricBoundaries as ee, Runtime as f, Schema as p, Stream as m } from "effect";
+import { and as h, eq as g, generateId as te, gte as ne, id as _, integer as v, json as y, queryFor as b, registerRetention as x, retentionTtlMsFromEnv as S, table as C, text as w, timestamp as T } from "@voltro/database";
+import { createHash as re } from "node:crypto";
+//#region src/tools.ts
+var ie = (e, t) => e(t), E = Symbol.for("@voltro/ai/Tool"), ae = (e) => ({
+	...e,
+	[E]: !0
+}), oe = (e) => typeof e == "object" && !!e && e[E] === !0, se = (e) => {
+	let t = e.default;
+	if (typeof t != "function") return 0;
+	let n = 0;
+	for (let r of Object.values(e)) oe(r) && r.execute === void 0 && r.handler === void 0 && (r.handler = t, n += 1);
+	return n;
+}, D = (e, t) => {
+	let n = t ?? f.defaultRuntime;
+	return f.runPromise(n)(e);
+}, O = (e, t, n) => n ? e.pipe(l.timeoutFail({
+	duration: `${n} millis`,
+	onTimeout: () => /* @__PURE__ */ Error(`tool '${t}' timed out`)
+})) : e, ce = (e, t, n) => {
+	if (e.execute) return D(O(ie(e.execute, t), e.name, e.timeoutMs), n.runtime);
+	if (e.handler) {
+		let r = e.handler, i = r(t, n);
+		return l.isEffect(i) ? D(O(i, e.name, e.timeoutMs), n.runtime) : i;
+	}
+	return Promise.reject(/* @__PURE__ */ Error(`tool '${e.name}' has neither execute nor handler`));
+}, le = (e) => {
+	let t = u.make(e);
+	return t && t.type === "object" ? t : {
+		type: "object",
+		properties: {},
+		additionalProperties: !1
+	};
+}, k = (e, t = { store: void 0 }) => {
+	let n = {};
+	for (let [r, i] of Object.entries(e)) n[i.name || r] = s({
+		description: i.description,
+		inputSchema: a(le(i.input)),
+		execute: async (e) => {
+			let n = await l.runPromise(p.decodeUnknown(i.input)(e)), r = await ce(i, n, t);
+			return i.output ? l.runPromise(p.decodeUnknown(i.output)(r)) : r;
+		}
+	});
+	return n;
+}, A = {
+	"claude-opus-4-8": {
+		inputPer1M: 15,
+		outputPer1M: 75
+	},
+	"claude-opus-4-6": {
+		inputPer1M: 15,
+		outputPer1M: 75
+	},
+	"claude-sonnet-4-6": {
+		inputPer1M: 3,
+		outputPer1M: 15
+	},
+	"claude-haiku-4-5": {
+		inputPer1M: 1,
+		outputPer1M: 5
+	},
+	"claude-haiku-4-5-20251001": {
+		inputPer1M: 1,
+		outputPer1M: 5
+	},
+	"gpt-5.5": {
+		inputPer1M: 5,
+		outputPer1M: 30
+	},
+	"gpt-4o": {
+		inputPer1M: 2.5,
+		outputPer1M: 10
+	},
+	"gpt-4o-mini": {
+		inputPer1M: .15,
+		outputPer1M: .6
+	},
+	"gemini-2.5-pro": {
+		inputPer1M: 1.25,
+		outputPer1M: 10
+	},
+	"gemini-2.5-flash": {
+		inputPer1M: .3,
+		outputPer1M: 2.5
+	},
+	"gemini-2.0-flash": {
+		inputPer1M: .1,
+		outputPer1M: .4
+	}
+}, j = {}, ue = (e) => {
+	let t = j;
+	return j = { ...e }, t;
+}, M = () => ({
+	...A,
+	...j
+}), N = (e) => {
+	let t = M();
+	if (t[e] !== void 0) return t[e];
+	let n = e.indexOf("/");
+	if (n > 0) {
+		let r = e.slice(n + 1);
+		if (t[r] !== void 0) return t[r];
+	}
+}, P = (e, t) => {
+	let n = e.inputTokens ?? 0, r = e.outputTokens ?? 0;
+	if (t.actualCostUsd !== void 0) {
+		let e = n + r, i = e > 0 ? n / e : 0;
+		return {
+			inputTokens: n,
+			outputTokens: r,
+			inputCostUsd: t.actualCostUsd * i,
+			outputCostUsd: t.actualCostUsd * (1 - i),
+			totalCostUsd: t.actualCostUsd,
+			costSource: "gateway"
+		};
+	}
+	let i = t.price ?? N(t.model) ?? {
+		inputPer1M: 0,
+		outputPer1M: 0
+	}, a = n / 1e6 * i.inputPer1M, o = r / 1e6 * i.outputPer1M;
+	return {
+		inputTokens: n,
+		outputTokens: r,
+		inputCostUsd: a,
+		outputCostUsd: o,
+		totalCostUsd: a + o,
+		costSource: "estimated"
+	};
+}, F = "_voltro_ai_usage", I = C(F, {
+	id: _({ prefix: "aiu" }),
+	tenantId: w().nullable(),
+	provider: w(),
+	model: w(),
+	operation: w(),
+	agent: w().nullable(),
+	inputTokens: v().default(0),
+	outputTokens: v().default(0),
+	costMicroUsd: v().default(0),
+	costSource: w().default("estimated"),
+	calledAt: T().default("now")
+}).index(["tenantId"]), de = (e) => {
+	let t = e?.gateway?.cost, n = typeof t == "number" ? t : typeof t == "string" ? Number(t) : NaN;
+	return Number.isFinite(n) ? n : void 0;
+}, fe = (e, t) => (x({
+	table: F,
+	timeColumn: "calledAt",
+	ttlMs: S(process.env.VOLTRO_AI_USAGE_TTL_HOURS, 24 * 365)
+}), l.promise(async () => {
+	let n = P(t.usage, {
+		model: t.model,
+		...t.price === void 0 ? {} : { price: t.price },
+		...t.actualCostUsd === void 0 ? {} : { actualCostUsd: t.actualCostUsd }
+	});
+	return await e.insert(F, {
+		tenantId: t.tenantId ?? null,
+		provider: t.provider,
+		model: t.model,
+		operation: t.operation,
+		agent: t.agent ?? null,
+		inputTokens: n.inputTokens,
+		outputTokens: n.outputTokens,
+		costMicroUsd: Math.round(n.totalCostUsd * 1e6),
+		costSource: n.costSource,
+		calledAt: t.calledAt ?? /* @__PURE__ */ new Date()
+	}), n;
+})), pe = (e, t = {}) => l.promise(async () => {
+	let n = [...t.tenantId === void 0 ? [] : [g("tenantId", t.tenantId)], ...t.since === void 0 ? [] : [ne("calledAt", t.since)]], r = b(I), i = n.length === 0 ? r : r.where(n.length === 1 ? n[0] : h(...n));
+	return (await e.query(i.descriptor)).reduce((e, t) => e + (Number(t.costMicroUsd) || 0), 0) / 1e6;
+}), L = class extends p.TaggedError()("AiBudgetExceeded", {
+	tenantId: p.NullOr(p.String),
+	limitUsd: p.Number,
+	spentUsd: p.Number,
+	attemptedUsd: p.Number
+}) {}, R = "_voltro_ai_budget", z = C(R, {
+	id: _({ prefix: "aib" }),
+	bucket: w().maxLength(191),
+	tenantId: w().nullable(),
+	reservedMicroUsd: v().default(0),
+	updatedAt: T().default("now")
+}).unique("byAiBudgetBucket", ["bucket"]), me = (e, t) => `${e ?? "global"}::${t === void 0 ? "all" : String(t.getTime())}`, B = 32, he = (e, t) => l.gen(function* () {
+	x({
+		table: R,
+		timeColumn: "updatedAt",
+		ttlMs: S(process.env.VOLTRO_AI_USAGE_TTL_HOURS, 24 * 365)
+	});
+	let n = t.addUsd ?? 0, r = Math.round(n * 1e6), i = Math.round(t.limitUsd * 1e6), a = me(t.tenantId, t.since), o = t.tenantId ?? null, s = (e) => new L({
+		tenantId: o,
+		limitUsd: t.limitUsd,
+		spentUsd: e / 1e6,
+		attemptedUsd: n
+	}), c = yield* l.promise(async () => {
+		for (let t = 0; t < B; t++) {
+			t > 0 && await new Promise((e) => setTimeout(e, t));
+			let n = (await e.query(b(z).where(g("bucket", a)).descriptor))[0], s = n && Number(n.reservedMicroUsd) || 0;
+			if (s + r > i) return {
+				ok: !1,
+				reservedMicro: s
+			};
+			if (r === 0) return { ok: !0 };
+			if (n) {
+				if (await e.updateMany("_voltro_ai_budget", {
+					reservedMicroUsd: s + r,
+					updatedAt: /* @__PURE__ */ new Date()
+				}, { where: h(g("id", String(n.id)), g("reservedMicroUsd", s)) }) === 1) return { ok: !0 };
+			} else {
+				let t = te({
+					kind: "typeid",
+					prefix: "aib"
+				}, R), n = await e.insertIgnore(R, {
+					id: t,
+					bucket: a,
+					tenantId: o,
+					reservedMicroUsd: r,
+					updatedAt: /* @__PURE__ */ new Date()
+				}, { conflictColumns: ["bucket"] });
+				if (String(n.id) === t) return { ok: !0 };
+			}
+		}
+		throw Error(`requireAiBudget: contention on bucket '${a}' exceeded ${B} attempts`);
+	});
+	if (!c.ok) return yield* l.fail(s(c.reservedMicro));
+}), ge = ee.exponential({
+	start: .001,
+	factor: 2,
+	count: 16
+}), V = d.counter("voltro_ai_calls_total", { description: "Total AI provider calls (generate + stream)." }), H = d.counter("voltro_ai_call_errors_total", { description: "AI provider calls that errored." }), U = d.histogram("voltro_ai_call_duration_seconds", ge, "AI provider call duration in seconds."), _e = d.counter("voltro_ai_input_tokens_total", {
+	description: "Total prompt (input) tokens across AI calls.",
+	incremental: !0
+}), ve = d.counter("voltro_ai_output_tokens_total", {
+	description: "Total completion (output) tokens across AI calls.",
+	incremental: !0
+}), ye = d.counter("voltro_ai_cost_microusd_total", {
+	description: "Total estimated AI spend in micro-USD (USD × 1e6).",
+	incremental: !0
+}), W = (e, t) => ({
+	provider: e.name,
+	model: e.model ?? e.name,
+	operation: t
+}), G = (e, t, n) => e.pipe(l.tagMetrics("provider", t.provider), l.tagMetrics("model", t.model), l.tagMetrics("operation", t.operation), l.tagMetrics("status", n)), K = (e, t) => {
+	let n = t.inputTokens ?? 0, r = t.outputTokens ?? 0, i = P({
+		inputTokens: n,
+		outputTokens: r
+	}, { model: e.model }), a = Math.round(i.totalCostUsd * 1e6);
+	return l.all([
+		G(d.incrementBy(_e, n), e, "ok"),
+		G(d.incrementBy(ve, r), e, "ok"),
+		G(d.incrementBy(ye, a), e, "ok")
+	], { discard: !0 });
+}, q = (e, t, n) => l.gen(function* () {
+	let r = yield* l.clockWith((e) => e.currentTimeMillis);
+	return yield* n.pipe(l.tap((n) => K(e, t(n))), l.tap(() => G(d.increment(V), e, "ok").pipe(l.zipRight(l.clockWith((e) => e.currentTimeMillis).pipe(l.flatMap((t) => G(d.update(U, (t - r) / 1e3), e, "ok")))))), l.tapError(() => G(d.increment(V), e, "error").pipe(l.zipRight(G(d.increment(H), e, "error")), l.zipRight(l.clockWith((e) => e.currentTimeMillis).pipe(l.flatMap((t) => G(d.update(U, (t - r) / 1e3), e, "error")))))), l.withSpan("voltro.ai.call", { attributes: {
+		"ai.provider": e.provider,
+		"ai.model": e.model,
+		"ai.operation": e.operation
+	} }));
+}), be = (e, t) => {
+	let n = Date.now();
+	return t.pipe(m.tap((t) => t._tag === "done" ? G(d.increment(V), e, "ok").pipe(l.zipRight(G(d.update(U, (Date.now() - n) / 1e3), e, "ok")), l.zipRight(K(e, t.usage ?? {}))) : t._tag === "error" ? G(d.increment(V), e, "error").pipe(l.zipRight(G(d.increment(H), e, "error")), l.zipRight(G(d.update(U, (Date.now() - n) / 1e3), e, "error"))) : l.void), m.withSpan("voltro.ai.call", { attributes: {
+		"ai.provider": e.provider,
+		"ai.model": e.model,
+		"ai.operation": e.operation
+	} }));
+}, J = class extends p.TaggedError()("AiError", {
+	reason: p.Literal("generation", "decode"),
+	message: p.String
+}) {}, Y = (e, t, n) => {
+	let r = [e, ...t ?? []], i = (e) => n(r[e]).pipe(l.catchTag("AiError", (t) => t.reason === "generation" && e + 1 < r.length ? i(e + 1) : l.fail(t)));
+	return i(0);
+}, X = (n) => Y(n.provider ?? t(), n.fallbacks, (t) => q(W(t, "generateText"), (e) => e.usage, l.tryPromise({
+	try: async () => {
+		let r = await i({
+			model: e(t),
+			prompt: n.prompt,
+			...n.system === void 0 ? {} : { system: n.system },
+			...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
+			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
+		});
+		return {
+			text: r.text,
+			usage: {
+				inputTokens: r.usage.inputTokens,
+				outputTokens: r.usage.outputTokens,
+				totalTokens: r.usage.totalTokens
+			},
+			providerMetadata: r.providerMetadata
+		};
+	},
+	catch: (e) => new J({
+		reason: "generation",
+		message: String(e?.message ?? e)
+	})
+}))), Z = (e, t) => l.gen(function* () {
+	let n = u.make(e);
+	if (n?.type !== "object") {
+		let e = n?.anyOf === void 0 ? `type:${JSON.stringify(n?.type ?? null)}` : "anyOf (e.g. Schema.Struct({}) or a union)";
+		return yield* l.fail(new J({
+			reason: "generation",
+			message: `${t}: the output schema must render as a top-level JSON object (type:"object"), but got ${e}. Use a Schema.Struct({ ... }) as the output schema — wrap arrays/scalars in a field (e.g. Schema.Struct({ items: Schema.Array(...) })).`
+		}));
+	}
+	return n;
+}), xe = (e, t) => Z(e, t), Q = (n) => Y(n.provider ?? t(), n.fallbacks, (t) => q(W(t, "generateObject"), (e) => e.usage, l.gen(function* () {
+	let i = e(t), o = yield* l.tryPromise({
+		try: () => r({
+			model: i,
+			schema: a(n.jsonSchema),
+			prompt: n.prompt,
+			...n.system === void 0 ? {} : { system: n.system },
+			...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
+			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
+		}),
+		catch: (e) => new J({
+			reason: "generation",
+			message: String(e?.message ?? e)
+		})
+	});
+	return {
+		object: o.object,
+		usage: {
+			inputTokens: o.usage.inputTokens,
+			outputTokens: o.usage.outputTokens,
+			totalTokens: o.usage.totalTokens
+		},
+		providerMetadata: o.providerMetadata
+	};
+}))), Se = (e) => l.gen(function* () {
+	let t = yield* Q({
+		jsonSchema: yield* Z(e.schema, "generateObject"),
+		prompt: e.prompt,
+		...e.system === void 0 ? {} : { system: e.system },
+		...e.provider === void 0 ? {} : { provider: e.provider },
+		...e.fallbacks === void 0 ? {} : { fallbacks: e.fallbacks },
+		...e.maxTokens === void 0 ? {} : { maxTokens: e.maxTokens },
+		...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+	}), n = yield* p.decodeUnknown(e.schema)(t.object).pipe(l.mapError((e) => new J({
+		reason: "decode",
+		message: `model output failed schema decode: ${String(e)}`
+	})));
+	return {
+		...t,
+		object: n
+	};
+}), Ce = (r) => l.gen(function* () {
+	let s = yield* l.runtime(), c = yield* Z(r.schema, "generateObjectWithTools"), u = {
+		store: void 0,
+		...r.toolContext,
+		runtime: s
+	}, d = k(r.tools, u);
+	return yield* Y(r.provider ?? t(), r.fallbacks, (t) => q(W(t, "generateObjectWithTools"), (e) => e.usage, l.gen(function* () {
+		let s = e(t), u = yield* l.tryPromise({
+			try: () => i({
+				model: s,
+				prompt: r.prompt,
+				tools: d,
+				stopWhen: o(r.maxSteps ?? 8),
+				output: n.object({ schema: a(c) }),
+				...r.system === void 0 ? {} : { system: r.system },
+				...r.maxTokens === void 0 ? {} : { maxOutputTokens: r.maxTokens },
+				...r.providerOptions === void 0 ? {} : { providerOptions: r.providerOptions }
+			}),
+			catch: (e) => new J({
+				reason: "generation",
+				message: String(e?.message ?? e)
+			})
+		});
+		return {
+			object: yield* p.decodeUnknown(r.schema)(u.output).pipe(l.mapError((e) => new J({
+				reason: "decode",
+				message: `model output failed schema decode: ${String(e)}`
+			}))),
+			usage: {
+				inputTokens: u.usage.inputTokens,
+				outputTokens: u.usage.outputTokens,
+				totalTokens: u.usage.totalTokens
+			},
+			providerMetadata: u.providerMetadata
+		};
+	})));
+}), we = class extends c.Tag("@voltro/AiService")() {}, Te = {
+	generateText: (e) => X(e),
+	generateObject: (e) => Se(e)
+}, Ee = "_voltro_ai_inferences", De = C(Ee, {
+	id: _({ prefix: "aiinf" }),
+	tenantId: w().nullable(),
+	executionId: w(),
+	workflowName: w(),
+	runId: w().nullable(),
+	signalName: w(),
+	stepName: w(),
+	status: w().oneOf([
+		"pending",
+		"running",
+		"succeeded",
+		"failed"
+	]).default("pending"),
+	kind: w().oneOf(["text", "object"]).default("text"),
+	provider: w(),
+	model: w(),
+	prompt: w(),
+	system: w().nullable(),
+	maxTokens: v().nullable(),
+	providerOptions: y().nullable(),
+	jsonSchema: y().nullable(),
+	result: y().nullable(),
+	errorMessage: w().nullable(),
+	attempts: v().default(0),
+	leaseUntil: T().nullable(),
+	claimedBy: w().nullable(),
+	enqueuedAt: T().default("now"),
+	startedAt: T().nullable(),
+	completedAt: T().nullable(),
+	durationMs: v().nullable()
+}).index(["status", "enqueuedAt"]).index(["status", "leaseUntil"]).index(["executionId"]).index(["tenantId"]), Oe = 5 * 6e4, ke = 4, Ae = (e, t) => {
+	let n = typeof e == "number" ? e : Number(e);
+	return Number.isFinite(n) ? Math.trunc(n) : t;
+}, $ = (e) => e == null ? null : String(e), je = (e) => ({
+	id: String(e.id),
+	tenantId: $(e.tenantId),
+	executionId: String(e.executionId),
+	workflowName: String(e.workflowName),
+	runId: $(e.runId),
+	signalName: String(e.signalName),
+	stepName: String(e.stepName),
+	kind: e.kind === "object" ? "object" : "text",
+	provider: String(e.provider),
+	model: String(e.model),
+	prompt: String(e.prompt ?? ""),
+	system: $(e.system),
+	maxTokens: e.maxTokens === null || e.maxTokens === void 0 ? null : Ae(e.maxTokens, 0),
+	providerOptions: e.providerOptions ?? null,
+	jsonSchema: e.jsonSchema ?? null,
+	attempts: Ae(e.attempts, 0)
+}), Me = (e) => {
+	let t = e.attempts + 1;
+	return t >= e.maxAttempts ? {
+		kind: "give-up",
+		attempts: t
+	} : {
+		kind: "retry",
+		attempts: t,
+		delayMs: e.retryAfterMs !== void 0 && e.retryAfterMs > 0 ? e.retryAfterMs : Math.min(e.baseDelayMs * 2 ** (t - 1), e.maxDelayMs)
+	};
+}, Ne = (e) => `voltro.ai.infer:${e}`, Pe = (e, t) => `aiinf_${re("sha256").update(`${e}\u0000${t}`).digest("hex").slice(0, 40)}`;
+//#endregion
+export { N as A, A as C, P as D, I as E, ae as F, oe as I, se as L, he as M, ue as N, de as O, k as P, L as S, pe as T, W as _, Me as a, R as b, je as c, Te as d, Se as f, xe as g, X as h, De as i, fe as j, M as k, J as l, Ce as m, ke as n, Ne as o, Q as p, Oe as r, Pe as s, Ee as t, we as u, be as v, z as w, F as x, q as y };