@bivelio/savings-layer 0.4.26 → 0.4.29

package/dist/index.d.ts

@@ -311,7 +311,7 @@ declare function precioDeLista(model: string, at?: Date | number): CostModel | u
  * on the customer's machine. `test/version-declarada.test.ts` ties this
  * constant to package.json so the two can never drift.
  */
-declare const BV_SALA_VERSION = "0.4.26";
+declare const BV_SALA_VERSION = "0.4.29";
 
 /**
  * Environment-agnostic UUID generation.
@@ -457,19 +457,38 @@ interface ChatMessage {
      * The provider tool loop needs it: OpenAI's Chat Completions rejects a
      * `role: "tool"` message that does not carry `tool_call_id` and does not follow
      * an assistant turn with `tool_calls` (a 400). Pass the `id` from the
-     * {@link ToolCallRequest} the model asked for. Ignored by providers that model
-     * tool results differently.
+     * {@link ToolCallRequest} the model asked for.
+     *
+     * Anthropic's Messages API needs it too: with it, the Anthropic adapter sends
+     * the result as a `tool_result` block paired with the call's `tool_use` block.
+     * Without it (or without the ids in the assistant turn's `toolCalls`), the
+     * exchange is sent as plain text and the adapter warns once.
      */
     toolCallId?: string;
+    /**
+     * For `role: "tool"`: the tool failed, and `content` says how.
+     *
+     * The Anthropic adapter sends it as `is_error: true` on the `tool_result`, so
+     * the model reads the content as an error rather than as data. Providers
+     * without such a flag (OpenAI's Chat Completions) send the content unchanged,
+     * so write the error into `content` either way.
+     */
+    isError?: boolean;
     /**
      * For `role: "assistant"`: the tool calls the model requested on that turn.
      *
      * Send it back with the assistant turn when you continue a tool conversation,
      * so the following `role: "tool"` messages are valid. These are the same
-     * {@link ToolCallRequest}s that came back in {@link SavingsResult.toolCalls}.
+     * {@link ToolCallRequest}s that came back in {@link SavingsResult.toolCalls},
+     * ids included.
      */
     toolCalls?: ToolCallRequest[];
-    /** Opaque provider metadata; never used for the exact-cache key. */
+    /**
+     * Opaque provider metadata; never used for the exact-cache key. The one
+     * exception is `tool_call_id` on a `role: "tool"` message: the OpenAI
+     * adapter still sends it as the call id when `toolCallId` is absent,
+     * so the key reads it as that same field.
+     */
     metadata?: Record<string, unknown>;
 }
 /** How a request is classified by the execution router. */
@@ -1107,6 +1126,14 @@ interface SavingsResult<T = unknown> {
      * (caché/resolver), que solo sirven respuestas finales completas.
      */
     finishReason?: string;
+    /**
+     * The refusal message the model returned instead of an answer (OpenAI
+     * Structured Outputs: `message.refusal`, with `finish_reason: "stop"` and no
+     * content). When present, the request was refused: `text` is empty and no
+     * output contract was met. It is measured like any call, and it is never
+     * repaired, cached or replayed.
+     */
+    refusal?: string;
     report: SavingsReport;
 }
 /** One event of a streamed answer. */
@@ -1281,8 +1308,12 @@ interface QualityRepairCheck {
      *   sending the request without the layer (opt-in, `skipRepairAboveBaseline`).
      * - `decode_time_constraint`: generation was already constrained, so a retry
      *   would not change the outcome.
+     * - `provider_refusal`: the provider refused the request (`finishReason:
+     *   "content_filter"`, Anthropic's `stop_reason: "refusal"`, or OpenAI's
+     *   `message.refusal`); a repair on the same model would be refused, and may
+     *   be billed, again.
      */
-    skipReason?: "output_truncated" | "schema_defect" | "would_cost_more_than_no_layer" | "decode_time_constraint";
+    skipReason?: "output_truncated" | "schema_defect" | "would_cost_more_than_no_layer" | "decode_time_constraint" | "provider_refusal";
 }
 /** Whether the provider stopped because it reached the output limit. */
 interface QualityTruncationCheck {
@@ -1832,6 +1863,25 @@ interface CanonicalRequest {
     normalizedText: string;
     systemPromptHash: string;
     toolsetHash: string;
+    /**
+     * Hash of the tool-loop fields of the conversation, the part of each turn
+     * that `normalizedText` does not carry: the calls on an assistant turn (name,
+     * arguments and id) and, on a `role: "tool"` message, the call it answers,
+     * its error flag and its name. Part of the exact key, the in-flight key and
+     * the semantic filter, so two conversations that differ only there never
+     * share an answer.
+     *
+     * Call ids count by what they pair, not by their value: results matched to
+     * other calls give another hash, while the same conversation with its ids
+     * consistently renamed (a new run of the same agent loop gets fresh ids from
+     * the provider) keeps it.
+     *
+     * Absent when the request carries none of them: the keys are then built from
+     * exactly the same material as before this field existed. A field that is
+     * present but empty or false (`toolCalls: []`, `isError: false`, `name: ""`)
+     * reaches the provider exactly as an absent one, and counts as absent.
+     */
+    toolTurnsHash?: string;
     /** Hash of the structured payload; part of the cache key so different data never collides. */
     structuredDataHash: string;
     schemaVersion: string;
@@ -1878,6 +1928,15 @@ declare function stableStringify(value: unknown): string;
 declare function modelFamilyOf(model: string): string;
 declare function canonicalize(request: SavingsRequest): CanonicalRequest;
 
+/**
+ * Thrown by `importCalibration` / `exportCalibration` when a calibration state
+ * cannot be used as-is: another embedder, verifier, judge or configuration, a
+ * newer format, or numbers that do not add up.
+ */
+declare class CalibrationStateError extends Error {
+    constructor(message: string);
+}
+
 /**
  * Decides whether a stored answer would have been right for a new request,
  * given the answer your provider just produced for it. `true` = the stored
@@ -1929,6 +1988,70 @@ interface SemanticCalibrationOptions {
     perRecordFloors?: boolean;
     /** Source of randomness for audits, in [0, 1). Default `Math.random`. */
     random?: () => number;
+    /**
+     * Names your `judge` (e.g. `"support-answers-v2"`), so that a saved
+     * calibration is only ever restored under the same definition of a wrong
+     * reuse. Required to export a calibration when you pass your own judge.
+     */
+    judgeId?: string;
+}
+
+/**
+ * A calibration saved with `exportCalibration`, to be restored with
+ * `importCalibration` after a restart. Plain JSON: store it as is.
+ *
+ * It carries the whole experiment, not a summary: both halves of the labels
+ * (the half that proposes floors and the half kept apart to test them), how
+ * many tests have already spent the confidence budget, and the floor in force.
+ * Restoring it continues the same calibration, so the bound keeps holding
+ * across restarts; it never starts the budget over.
+ *
+ * It describes itself: who scored the requests (`embedder`, `verifier`), who
+ * judged the answers (`judge`) and the numbers of the test (`config`). A cache
+ * with any of them different refuses it.
+ */
+interface SemanticCalibrationState {
+    format: "bvsala.semantic-calibration";
+    /** Format and method version. A newer SDK may read an older state; never the other way round. */
+    version: 1;
+    /** `built_in:<version>` for the embedder shipped with the SDK, `custom:<embedderId>` for yours. */
+    embedder: string;
+    /** `null` without a verifier, `custom:<verifierId>` with yours. */
+    verifier: string | null;
+    /** `built_in:<version>` for {@link sameAnswer}, `custom:<judgeId>` for yours. */
+    judge: string;
+    config: {
+        maxErrorRate: number;
+        confidence: number;
+        /** Only candidates at or above it are labeled, so it is part of what the labels mean. */
+        similarityThreshold: number;
+        /** Candidate floors are tested on a fixed grid of `grid + 1` values in [0, 1]. */
+        grid: number;
+    };
+    /** One calibration per traffic segment; a segment with no state yet is absent. */
+    segments: {
+        single_shot?: SemanticCalibrationSegmentState;
+        agentic?: SemanticCalibrationSegmentState;
+    };
+}
+/** The calibration of one traffic segment. Grid indices are scores times `grid`. */
+interface SemanticCalibrationSegmentState {
+    /** Labels recorded. They alternate by arrival between the two halves below. */
+    labeled: number;
+    /** Tests already run. It follows from `labeled`, and a state where it does not is refused. */
+    looks: number;
+    /** The floor the next test starts from (a grid index), or null. */
+    previousCandidate: number | null;
+    /** The floor in force and the held-out labels that certified it, or null. */
+    certified: {
+        floorIndex: number;
+        heldOut: number;
+        wrong: number;
+    } | null;
+    /** The half that proposes floors: `[gridIndex, labels, wrong]` where there are labels. */
+    selection: [number, number, number][];
+    /** The half kept apart to test them, in the same shape. */
+    heldOut: [number, number, number][];
 }
 /** One labeled reuse decision: the score the cache would have served on, and whether the stored answer was right. */
 interface CalibrationObservation {
@@ -2032,6 +2155,13 @@ interface SemanticRecord {
     safetyPolicyVersion: string;
     /** Hash of the declared tool catalog (ids, risk and parameters). */
     toolsetHash: string;
+    /**
+     * Hash of the tool-loop fields of the conversation (tool calls, the call each
+     * tool result answers, its error flag and name); absent when it had none.
+     * Compared for equality like the other identity fields: similarity is scored
+     * on text only, and never decides between two tool loops.
+     */
+    toolTurnsHash?: string;
     /** Sampling temperature of the request, or null when it set none. */
     temperature: number | null;
     /** Sampling seed of the request, or null when it set none. */
@@ -2100,7 +2230,16 @@ interface SemanticCacheOptions {
      * cache miss — never a wrong answer.
      */
     embed?: EmbeddingFn;
+    /**
+     * Names your `embed` function: model and version, e.g.
+     * `"text-embedding-3-small@1536"`. A saved calibration carries it, and one
+     * recorded under another embedder is refused on import. Required to export a
+     * calibration when you pass your own `embed`.
+     */
+    embedderId?: string;
     verify?: SemanticVerifier;
+    /** Names your `verify` function, like {@link embedderId}. Required to export a calibration when you pass one. */
+    verifierId?: string;
     /** Cosine similarity floor for a candidate to be considered at all. */
     similarityThreshold?: number;
     /** Verifier score floor for acceptance. */
@@ -2140,6 +2279,7 @@ declare class InMemorySemanticCache {
     private readonly embebedorPropio;
     private readonly calibracion;
     private readonly estados;
+    private readonly nombres;
     private readonly pendientes;
     private readonly aceptacionExplicita;
     constructor(opts?: SemanticCacheOptions);
@@ -2184,6 +2324,38 @@ declare class InMemorySemanticCache {
     settled(): Promise<void>;
     /** Where the calibration of a traffic segment stands, or null without calibration. */
     calibrationStatus(segment?: SemanticProbe["segment"]): SemanticCalibrationStatus | null;
+    /**
+     * The calibration's full state, to restore it with {@link importCalibration}
+     * in a later process: after a deploy, or in the next serverless invocation.
+     * Plain JSON. Call it after `await cache.settled()` so labels still being
+     * judged are included.
+     *
+     * The bound holds along one line of history: restore the latest state you
+     * saved and keep saving from where it went. Processes that continue the same
+     * state at the same time each spend its remaining confidence again. Each one
+     * keeps its own bound, but the chance that some process certifies a wrong
+     * floor grows with how many do it, and whichever saves last overwrites the
+     * labels of the others.
+     *
+     * Throws {@link CalibrationStateError} without calibration, and when you
+     * passed your own `embed`, `verify` or `judge` without naming it
+     * (`embedderId`, `verifierId`, `calibration.judgeId`).
+     */
+    exportCalibration(): SemanticCalibrationState;
+    /**
+     * Restore a state saved with {@link exportCalibration} (the object, or its
+     * JSON text), right after constructing the cache. The calibration continues
+     * exactly where it was: the same labels in the same two halves, the same
+     * tests already run and the same floor in force.
+     *
+     * All or nothing. Throws {@link CalibrationStateError}, and changes nothing,
+     * when the state comes from another embedder, verifier, judge or
+     * configuration, from a newer SDK, when its numbers do not add up, or when
+     * this cache has already recorded labels in a segment the state carries
+     * (merging two histories would count some labels twice).
+     */
+    importCalibration(state: SemanticCalibrationState | string | unknown): void;
+    private identidad;
     private estadoDe;
     store(c: CanonicalRequest, answer: CachedAnswer): Promise<void>;
     /** Drop every record for a knowledge epoch (invalidation on re-index). */
@@ -2191,6 +2363,55 @@ declare class InMemorySemanticCache {
     get size(): number;
 }
 
+/**
+ * Where the layer keeps what it learned about your prompt-cache reuse
+ * (`cacheBreakpoints: "auto"`), when it must outlive a single instance.
+ *
+ * By default that measurement lives in the layer's memory, which is enough for
+ * a long-running server. On serverless platforms, where an instance serves one
+ * request or lives a few minutes, memory is gone before the measurement means
+ * anything, so the layer never gets past its fixed placement. Pass a store
+ * shared by your instances (Redis, a KV) and every instance reads and extends
+ * the same measurement.
+ *
+ * What is stored, per traffic class (model, tools, output contract and system
+ * header): prompt-prefix fingerprints, the times they were sent, reuse counts
+ * and how many predicted cache reads the provider confirmed (with
+ * `cacheMissDiagnostics`, also the provider's response ids). Never prompt text
+ * or any other content. Values are opaque JSON strings, one key per traffic
+ * class, rewritten after each request of that class. A client that hands them
+ * back already parsed (as some Redis and KV clients do) works as well.
+ *
+ * The interface is deliberately minimal so any key-value store can back it.
+ * The layer awaits `get` before placing breakpoints and `get` + `set` after the
+ * provider answers, so a slow store adds its latency to those requests. A store
+ * that fails is treated as empty for that request, and nothing is written
+ * back: the placement falls back to what a fresh process would do.
+ */
+interface CacheBreakpointStore {
+    /** The value last `set` for `key` (the string, or the same JSON already parsed), or null/undefined when there is none. */
+    get(key: string): Promise<unknown>;
+    /** `ttlMs`: drop the value if no request of its traffic class rewrites it for this long. What it resolves to is ignored. */
+    set(key: string, value: string, ttlMs: number): Promise<unknown>;
+}
+/**
+ * In-memory {@link CacheBreakpointStore}, with TTL and a size bound. Useful for
+ * tests and for several layers in one process; across processes, back the
+ * interface with a shared store.
+ */
+declare class InMemoryCacheBreakpointStore implements CacheBreakpointStore {
+    private readonly entradas;
+    private readonly maxEntries;
+    private readonly now;
+    constructor(opts?: {
+        maxEntries?: number;
+        now?: () => number;
+    });
+    get(key: string): Promise<string | null>;
+    set(key: string, value: string, ttlMs: number): Promise<void>;
+    get size(): number;
+}
+
 /**
  * Singleflight (in-flight coalescing).
  *
@@ -2834,6 +3055,16 @@ interface ProviderRequest {
      * Absent: the adapter's fixed placement.
      */
     cachePlacement?: CachePlacement;
+    /**
+     * Ask the provider to compare this request's prompt prefix with the one of
+     * an earlier response and report where they diverged (Anthropic cache
+     * diagnostics, Claude API only). `previousMessageId: null` opts in without a
+     * comparison, so that a later request can be compared with this one.
+     * Adapters without the feature, or talking to another host, ignore it.
+     */
+    cacheDiagnostics?: {
+        previousMessageId: string | null;
+    };
 }
 /** See {@link ProviderRequest.toolSearch}. */
 interface ProviderToolSearch {
@@ -2850,6 +3081,15 @@ interface ProviderToolSearchUsage {
     loaded: string[];
     /** Searches that came back as an error result instead of references. */
     errors: number;
+    /**
+     * Search results from earlier turns that the adapter sent back in this request's history,
+     * so the model could call the tools they loaded without searching again: how many results,
+     * and which tools. Absent when none were sent back.
+     */
+    resent?: {
+        results: number;
+        tools: string[];
+    };
 }
 /**
  * Breakpoint placement decided by the layer from the reuse it measured on your
@@ -2938,6 +3178,17 @@ interface ProviderResponse {
      * podía distinguirlo y hacía exactamente eso.
      */
     finishReason?: "stop" | "length" | "content_filter" | "tool_calls" | (string & {});
+    /**
+     * The refusal message the model returned INSTEAD of an answer, exactly as the
+     * provider wrote it: OpenAI's `message.refusal` (Chat Completions with
+     * Structured Outputs), which arrives with `finish_reason: "stop"` and no
+     * `content`. Present only when non-empty.
+     *
+     * The layer treats it like `finishReason: "content_filter"`: the call is
+     * measured, and the answer is never repaired against its contract, cached, or
+     * replayed by reference. `text` and `finishReason` stay what the provider said.
+     */
+    refusal?: string;
     /**
      * Llamadas a herramienta que pidió el modelo, tal cual las devolvió el
      * proveedor.
@@ -2986,6 +3237,22 @@ interface ProviderResponse {
      * read what the search did. The layer only reports it; it credits nothing for it.
      */
     toolSearch?: ProviderToolSearchUsage;
+    /** The provider's own id for this response (Anthropic `id`), when it gave one. */
+    providerMessageId?: string;
+    /**
+     * The provider's answer to {@link ProviderRequest.cacheDiagnostics}, when the
+     * request asked for a comparison. `missReason` is the provider's
+     * `cache_miss_reason.type` (`model_changed`, `system_changed`,
+     * `tools_changed`, `messages_changed`, `previous_message_not_found`,
+     * `unavailable`), `"none"` when it compared and found no divergence, or
+     * `"pending"` when the comparison had not finished. `missedInputTokens` is the
+     * provider's estimate of the input after the divergence: a magnitude, not a
+     * billing number.
+     */
+    cacheDiagnostics?: {
+        missReason: string;
+        missedInputTokens?: number;
+    };
     raw?: unknown;
 }
 /** One event of a provider-level stream. */
@@ -3277,6 +3544,35 @@ interface SavingsLayerConfig {
      * causes, one-hour writes at 2x included, is charged to `cost.actual`.
      */
     cacheBreakpoints?: "fixed" | "auto";
+    /**
+     * Where `cacheBreakpoints: "auto"` keeps its reuse measurement when it must
+     * outlive a single instance. Optional: without it the measurement lives in this
+     * layer's memory, exactly as before.
+     *
+     * On serverless platforms (one instance per request, or instances that live
+     * a few minutes) memory never gathers enough samples, so `"auto"` stays on
+     * the fixed placement. Pass a store your instances share and each one reads
+     * and extends the same measurement: what one instance learned, the next one
+     * uses. Only prompt-prefix fingerprints, timestamps and counts are stored,
+     * never content. See {@link CacheBreakpointStore}.
+     */
+    cacheBreakpointStore?: CacheBreakpointStore;
+    /**
+     * With `cacheBreakpoints: "auto"` on Anthropic's own API: when the layer
+     * expects a request to read the cache entry an earlier response wrote, ask
+     * Anthropic to compare the two requests (cache diagnostics), and when that
+     * read does not arrive, publish the provider's `cache_miss_reason` in the
+     * `prefix-cache-placement` trace (`system_changed`, `tools_changed`,
+     * `messages_changed`…, or `none` when the requests matched and the entry was
+     * simply gone). Off by default.
+     *
+     * Each request that opts in makes Anthropic keep a fingerprint of it for a
+     * short time (hashes and token-count estimates, never prompt text), scoped to
+     * your organization and workspace; check it against your data-retention
+     * terms before switching this on. Not sent through Bedrock, Vertex or any
+     * other host. It costs nothing and changes no bill.
+     */
+    cacheMissDiagnostics?: boolean;
     /**
      * Signed BiVelio license token (from your paid subscription). This is the
      * primary way to license the SDK: `createSavingsLayer` verifies it offline
@@ -3479,7 +3775,8 @@ interface SavingsLayerConfig {
      * revocation list going stale. Nothing here throws and nothing here changes
      * what the layer does — turning it on only removes the console noise, so set
      * it when you have your own observability. Applies to the pipeline, the usage
-     * emitter and the revocation poller alike.
+     * emitter, the revocation poller and the provider adapters shipped in this
+     * package alike.
      */
     silenceWarnings?: boolean;
     /**
@@ -3859,7 +4156,8 @@ declare const registrarTokenizadorExacto: typeof registerExactTokenizer;
  *
  * A hash of the user text alone is NOT a safe cache key. Two requests may share
  * text but differ in tenant, permission scope, model version, system prompt,
- * schema, toolset, locale, sampling, knowledge epoch or safety policy — any of
+ * schema, toolset, tool-loop fields (tool calls, the call a tool result answers,
+ * its error flag), locale, sampling, knowledge epoch or safety policy — any of
  * which can change the correct answer. The exact key includes all of them.
  *
  * `servedModel` CARRIES THE MODEL THAT WILL ACTUALLY ANSWER, not the one the
@@ -4395,6 +4693,7 @@ declare class AnthropicProvider implements Provider {
     readonly id: string;
     private readonly opts;
     private readonly fetchImpl;
+    private readonly busquedas;
     constructor(opts?: AnthropicOptions);
     /** The id Anthropic's API sees: bare at `api.anthropic.com`, as written elsewhere. */
     private modeloEnElCable;
@@ -4411,8 +4710,12 @@ declare class AnthropicProvider implements Provider {
     generateStream(req: ProviderRequest, options?: ProviderStreamOptions): AsyncGenerator<ProviderStreamEvent, void, undefined>;
     private enviar;
     private construirCuerpo;
+    private reenvioDeBusquedas;
+    private recordarBusquedas;
+    private avisar;
     private marcarCache;
     private leer;
+    private diagnosticoDe;
     private leerMensaje;
 }
 
@@ -4652,7 +4955,6 @@ declare class RevocationPoller {
     private consecutiveFailures;
     private stopped;
     private started;
-    private warnedNoFetch;
     /** ETag of the last 200; sent as If-None-Match so an unchanged list 304s (cheap no-op). */
     private lastEtag;
     /** 304 seguidos; al llegar al tope se fuerza una descarga firmada. */
@@ -5290,4 +5592,4 @@ declare function toCsv(rows: readonly Readonly<Record<string, CsvValue>>[], opti
     columns?: readonly string[];
 }): string;
 
-export { type AnswerJudge, type AnswerSource, type AnthropicOptions, AnthropicProvider, BV_SALA_VERSION, type BatchCounts, type BatchItemResult, type BatchJob, type BatchJobStatus, type BatchRequest, type BatchResults, BatchSubmitError, type BatchSubmitOptions, type BatchSummary, type BivelioHandoff, CATALOGO, COMPACTION_TRANSFORMS, COMPACT_ROWS_INSTRUCTION, type CacheOutcome, type CachePlacement, type CacheTier, type CacheWriteReport, type CachedAnswer, type CalibrationObservation, type CanonicalRequest, type ChatMessage, type CompactRowsDecodeOptions, type CompactRowsOptions, type CompactVerification, type CompactionTransform, type CompressionEconomics, type CompressionPolicyInput, type CompressionRequest, type CompressionResult, type Compressor, type ConstraintBackend, type ConstraintBackendId, type ConstraintPlan, type ConversationState, type CostBasis, type CostModel, type CsvValue, DEFAULT_COST_MODEL, DEFAULT_MAX_STALENESS_SECONDS, DEFAULT_MIN_SAVED_CHARS, type DeferredExecutionReport, type DeterministicResolver, EMBEDDED_PUBLIC_KEY, type EmbeddingFn, type EvidenceDoc, type ExactCache, ExactTokenizer, type ExecutionClass, ExecutionRouter, type FocusExportOptions, type FocusRow, type FormatChoice, FormatSelector, type FormatSelectorOptions, type HandoffFact, HeuristicEstimator, InMemoryExactCache, InMemorySemanticCache, InMemoryToolSetFreezer, type JsonSchema, type LangfuseGenerationUpdate, LicenseError, type LicenseGate, type LicenseKeyShape, type LicenseKeyset, type LicenseStatus, type LicenseTokenPayload, type LocalClassifier, MODELOS_POR_CODIFICACION, MODELS_BY_ENCODING, type MemoryDelta, type MessageRole, MockProvider, type MockProviderOptions, type ModelCascadeMode, type ModelCascadeReport, type ModelCascadeSettings, type OpenAICompatibleOptions, OpenAICompatibleProvider, type OutputContract, PRECIOS_AL_DIA, type Policy, type PriceLeverSaving, type PriceLeverSettings, type ProtectedSpan, type Provider, type ProviderRequest, type ProviderResponse, type ProviderStreamEvent, type ProviderStreamOptions, type ProviderToolSearch, type ProviderToolSearchUsage, type ProviderUsage, QUALITY_STATUSES, type QualityCacheMatch, type QualityCascadeCheck, type QualityLedger, type QualityRepairCheck, type QualitySchemaCheck, type QualityStatus, type QualityTruncationCheck, REFERENCE_MAX_AGE_MS, REFERENCE_MESSAGE_ALLOWANCE, REFERENCE_OUTPUT_SHARE, type ReferenceCredit, type ReferenceRejection, type RequestContext, type Reranker, type ResolvedDeps, type ResolvedGate, type ResolvedLicense, ResolverRegistry, type ResolverResult, type RetrievalOptions, type RetrievalOutcome, type RetrievalPlan, type Retriever, type RevocationListBody, type RevocationListVerification, type RevocationPollEvent, RevocationPoller, type RevocationPollerOptions, RevocationSnapshot, type RouteDecision, type SavingsAdapterLevers, type SavingsAdapterOptions, type SavingsAttributeOptions, type SavingsBatches, type SavingsCallOptions, type SavingsLayer, type SavingsLayerConfig, SavingsLedger, type SavingsRecord, type SavingsReport, type SavingsReportForExport, type SavingsRequest, type SavingsResult, type SavingsSeriesOptions, type SavingsSeriesRow, type SavingsSpanAttributes, type SavingsSpanOptions, type SavingsStream, type SavingsStreamEvent, type SavingsStreamOptions, type SavingsToggles, type SavingsTracingOptions, SchemaRegistry, type SemanticCacheOptions, type SemanticCalibrationBound, type SemanticCalibrationOptions, type SemanticCalibrationStatus, type SemanticHit, type SemanticLookup, type SemanticProbe, type SemanticRecord, type SemanticVerifier, type SerializationFormat, type Serializer, Singleflight, type SpanAttributeValue, type SpanLike, type SpanStartOptionsLike, StreamAbortedError, type StreamingReport, type TokenEstimator, TokenizerRegistry, type ToolCallRequest, type ToolDescriptor, type ToolOutputCompaction, type ToolOutputCompactionSettings, type ToolRisk, type ToolSearchMode, type ToolSearchOptions, type ToolSelection, ToolSelector, type ToolSelectorOptions, type ToolSetFreezer, TraceCollector, type TraceEvent, type TracerLike, type TracerProviderLike, UMBRAL_LOCAL, UsageEmitter, type UsageEmitterOptions, type ValidationError, type ValidationResult, type VerificationCheck, type VerificationExclusion, type VerificationOutcome, type VerificationReport, type VerificationSettings, type VerificationSide, type VerificationStratum, adaptiveK, arithmeticResolver, buildExactKey, buildInflightKey, calibrateSemanticFloor, canonicalize, compactRowsSerializer, compactToolOutput, compresionFiel, computeNextDelay, cosine, createCompactRowsSerializer, createDefaultResolvers, createHandoff, createMockProvider, createSavingsLayer, csvSerializer, decodeCompactRows, dedupe, defaultTokenizerRegistry, deriveSeatId, deriveStatus, detectCi, embebedorLocal, emptyState, encodeCompactRows, evalArithmetic, evaluatePolicy, executeRetrievalPlan, gateFromKey, heuristicEstimator, instrumentSavingsLayer, isKeyset, isStrictCompatible, jsonCompactSerializer, keyTokensMatch, licenseKeyShape, liveGateFromKey, mergeDelta, modelFamilyOf, multisetEqual, negationCount, noopCompressor, normalizeLicenseKey, normalizeText, numericTokens, precioDeLista, projectContext, providerNativeBackend, quitarPrefijoDeProveedor, randomId, recordSavingsOnSpan, registerExactTokenizer, registrarTokenizadorExacto, renderState, resolveConstraintBackend, resolveDependencies, resolveLicense, sameAnswer, savingsSpanAttributes, sha256, shouldCompress, stableStringify, toCsv, toFocusRows, toLangfuseGeneration, toSavingsSeries, uniformColumns, validateAgainstSchema, vectorLocal, verifyCompactSignature, verifyLicenseToken, verifyRevocationList, withProtectedSpans, words };
+export { type AnswerJudge, type AnswerSource, type AnthropicOptions, AnthropicProvider, BV_SALA_VERSION, type BatchCounts, type BatchItemResult, type BatchJob, type BatchJobStatus, type BatchRequest, type BatchResults, BatchSubmitError, type BatchSubmitOptions, type BatchSummary, type BivelioHandoff, CATALOGO, COMPACTION_TRANSFORMS, COMPACT_ROWS_INSTRUCTION, type CacheBreakpointStore, type CacheOutcome, type CachePlacement, type CacheTier, type CacheWriteReport, type CachedAnswer, type CalibrationObservation, CalibrationStateError, type CanonicalRequest, type ChatMessage, type CompactRowsDecodeOptions, type CompactRowsOptions, type CompactVerification, type CompactionTransform, type CompressionEconomics, type CompressionPolicyInput, type CompressionRequest, type CompressionResult, type Compressor, type ConstraintBackend, type ConstraintBackendId, type ConstraintPlan, type ConversationState, type CostBasis, type CostModel, type CsvValue, DEFAULT_COST_MODEL, DEFAULT_MAX_STALENESS_SECONDS, DEFAULT_MIN_SAVED_CHARS, type DeferredExecutionReport, type DeterministicResolver, EMBEDDED_PUBLIC_KEY, type EmbeddingFn, type EvidenceDoc, type ExactCache, ExactTokenizer, type ExecutionClass, ExecutionRouter, type FocusExportOptions, type FocusRow, type FormatChoice, FormatSelector, type FormatSelectorOptions, type HandoffFact, HeuristicEstimator, InMemoryCacheBreakpointStore, InMemoryExactCache, InMemorySemanticCache, InMemoryToolSetFreezer, type JsonSchema, type LangfuseGenerationUpdate, LicenseError, type LicenseGate, type LicenseKeyShape, type LicenseKeyset, type LicenseStatus, type LicenseTokenPayload, type LocalClassifier, MODELOS_POR_CODIFICACION, MODELS_BY_ENCODING, type MemoryDelta, type MessageRole, MockProvider, type MockProviderOptions, type ModelCascadeMode, type ModelCascadeReport, type ModelCascadeSettings, type OpenAICompatibleOptions, OpenAICompatibleProvider, type OutputContract, PRECIOS_AL_DIA, type Policy, type PriceLeverSaving, type PriceLeverSettings, type ProtectedSpan, type Provider, type ProviderRequest, type ProviderResponse, type ProviderStreamEvent, type ProviderStreamOptions, type ProviderToolSearch, type ProviderToolSearchUsage, type ProviderUsage, QUALITY_STATUSES, type QualityCacheMatch, type QualityCascadeCheck, type QualityLedger, type QualityRepairCheck, type QualitySchemaCheck, type QualityStatus, type QualityTruncationCheck, REFERENCE_MAX_AGE_MS, REFERENCE_MESSAGE_ALLOWANCE, REFERENCE_OUTPUT_SHARE, type ReferenceCredit, type ReferenceRejection, type RequestContext, type Reranker, type ResolvedDeps, type ResolvedGate, type ResolvedLicense, ResolverRegistry, type ResolverResult, type RetrievalOptions, type RetrievalOutcome, type RetrievalPlan, type Retriever, type RevocationListBody, type RevocationListVerification, type RevocationPollEvent, RevocationPoller, type RevocationPollerOptions, RevocationSnapshot, type RouteDecision, type SavingsAdapterLevers, type SavingsAdapterOptions, type SavingsAttributeOptions, type SavingsBatches, type SavingsCallOptions, type SavingsLayer, type SavingsLayerConfig, SavingsLedger, type SavingsRecord, type SavingsReport, type SavingsReportForExport, type SavingsRequest, type SavingsResult, type SavingsSeriesOptions, type SavingsSeriesRow, type SavingsSpanAttributes, type SavingsSpanOptions, type SavingsStream, type SavingsStreamEvent, type SavingsStreamOptions, type SavingsToggles, type SavingsTracingOptions, SchemaRegistry, type SemanticCacheOptions, type SemanticCalibrationBound, type SemanticCalibrationOptions, type SemanticCalibrationSegmentState, type SemanticCalibrationState, type SemanticCalibrationStatus, type SemanticHit, type SemanticLookup, type SemanticProbe, type SemanticRecord, type SemanticVerifier, type SerializationFormat, type Serializer, Singleflight, type SpanAttributeValue, type SpanLike, type SpanStartOptionsLike, StreamAbortedError, type StreamingReport, type TokenEstimator, TokenizerRegistry, type ToolCallRequest, type ToolDescriptor, type ToolOutputCompaction, type ToolOutputCompactionSettings, type ToolRisk, type ToolSearchMode, type ToolSearchOptions, type ToolSelection, ToolSelector, type ToolSelectorOptions, type ToolSetFreezer, TraceCollector, type TraceEvent, type TracerLike, type TracerProviderLike, UMBRAL_LOCAL, UsageEmitter, type UsageEmitterOptions, type ValidationError, type ValidationResult, type VerificationCheck, type VerificationExclusion, type VerificationOutcome, type VerificationReport, type VerificationSettings, type VerificationSide, type VerificationStratum, adaptiveK, arithmeticResolver, buildExactKey, buildInflightKey, calibrateSemanticFloor, canonicalize, compactRowsSerializer, compactToolOutput, compresionFiel, computeNextDelay, cosine, createCompactRowsSerializer, createDefaultResolvers, createHandoff, createMockProvider, createSavingsLayer, csvSerializer, decodeCompactRows, dedupe, defaultTokenizerRegistry, deriveSeatId, deriveStatus, detectCi, embebedorLocal, emptyState, encodeCompactRows, evalArithmetic, evaluatePolicy, executeRetrievalPlan, gateFromKey, heuristicEstimator, instrumentSavingsLayer, isKeyset, isStrictCompatible, jsonCompactSerializer, keyTokensMatch, licenseKeyShape, liveGateFromKey, mergeDelta, modelFamilyOf, multisetEqual, negationCount, noopCompressor, normalizeLicenseKey, normalizeText, numericTokens, precioDeLista, projectContext, providerNativeBackend, quitarPrefijoDeProveedor, randomId, recordSavingsOnSpan, registerExactTokenizer, registrarTokenizadorExacto, renderState, resolveConstraintBackend, resolveDependencies, resolveLicense, sameAnswer, savingsSpanAttributes, sha256, shouldCompress, stableStringify, toCsv, toFocusRows, toLangfuseGeneration, toSavingsSeries, uniformColumns, validateAgainstSchema, vectorLocal, verifyCompactSignature, verifyLicenseToken, verifyRevocationList, withProtectedSpans, words };