pareta 0.2.1 → 1.1.0

package/dist/index.d.cts

@@ -11,9 +11,8 @@ interface SSEEvent {
  * typed layer.
  *
  * Alias / D3 boundary (a BACKEND contract, the SDK adds no alias logic):
- * `Endpoint.model`, `EvalResult.modelId`, leaderboard names and
- * `run.candidateModels` are per-task PUBLIC aliases — real open-weights ids
- * never cross. Frontier (vendor) ids are in the clear.
+ * `EvalResult.modelId` and `run.candidateModels` are per-task PUBLIC aliases —
+ * real open-weights ids never cross. Frontier (vendor) ids are in the clear.
  */
 type Raw = Record<string, any>;
 /** Base: holds the raw payload, dict-style escape hatches. */
@@ -63,29 +62,6 @@ declare class ModelList extends BaseModel implements Iterable<Model> {
     get length(): number;
     [Symbol.iterator](): Iterator<Model>;
 }
-/**
- * A deployed endpoint. `id` (== name) is what you pass to
- * `chat.completions.create({ model })`. `model` is the per-task public alias.
- */
-declare class Endpoint extends BaseModel {
-    get id(): string | null;
-    get name(): string | null;
-    get model(): string | null;
-    get status(): string | null;
-    get task(): string | null;
-    get url(): string | null;
-    get isLive(): boolean;
-    /** Structured-extraction endpoints (contract / SEC / ICD …): the system prompt
-     *  the benchmark used. The proxy applies it automatically when you send no
-     *  `system` message (so quality matches the leaderboard by default); returned
-     *  here so you can read it, or override it with your own `system` message.
-     *  `null` when the task has no injectable prompt (see `promptScaffold`). */
-    get recommendedSystemPrompt(): string | null;
-    /** Fixed-label classification endpoints: a copy-and-customize `system` prompt
-     *  template. NEVER auto-applied — the categories are yours, not the benchmark's
-     *  — so fill them in and send it as your `system` message. `null` otherwise. */
-    get promptScaffold(): string | null;
-}
 declare class Task extends BaseModel {
     get id(): string | null;
     get defaultScorer(): string | null;
@@ -139,29 +115,12 @@ declare class EvalResult extends BaseModel {
     /** Per-item rows (idx/score/prediction/error); empty when not persisted. */
     get perItem(): EvalItemResult[];
 }
-declare class LeaderboardEntry extends BaseModel {
-    get name(): string | null;
-    get kind(): string | null;
-    get quality(): number | null;
-    get costPerRequestMicroUsd(): number | null;
-    get contextK(): number | null;
-    get runMode(): string | null;
-}
-/** Models ranked for a task. `recommended` is the deployable pick; `frontier` the baseline. */
-declare class Leaderboard extends BaseModel {
-    get taskId(): string | null;
-    get metric(): string | null;
-    get costUnit(): string | null;
-    get recommended(): string | null;
-    get models(): LeaderboardEntry[];
-    get frontier(): LeaderboardEntry | null;
-}
 /** A vendor frontier model you can evaluate against (from the eval pool). */
 declare class FrontierModel extends BaseModel {
     get id(): string | null;
     get vendor(): string | null;
     get vision(): boolean;
-    /** Only meaningful when a task was given: it's on that task's leaderboard. */
+    /** Only meaningful when a task was given: it has benchmark results on that task. */
     get benchmarked(): boolean;
 }
 /**
@@ -183,13 +142,37 @@ declare class EvalRun extends BaseModel {
     get cost(): string;
     get results(): EvalResult[];
 }
+/** One row of `Rerank.results` — a document's position + calibrated score. */
+declare class RerankResult extends BaseModel {
+    /** Position of this document in YOUR request's documents array. */
+    get index(): number;
+    /** Calibrated P(relevant) in (0, 1) — thresholdable, not just ordinal. */
+    get relevanceScore(): number;
+}
+/** Result of `pa.rerank(...)`: `.results` ordered most-relevant-first;
+ * `.pairs` is the number of documents scored (the metered unit). */
+declare class Rerank extends BaseModel {
+    get results(): RerankResult[];
+    get model(): string | null;
+    get pairs(): number | null;
+    /** Map the ranked indices back onto the documents you sent — best first. */
+    topDocuments(documents: string[]): string[];
+}
+/** Result of `pa.embeddings(...)`: unit-normalized vectors in input order
+ * (cosine similarity is a plain dot product); `.promptTokens` is metered. */
+declare class Embeddings extends BaseModel {
+    get vectors(): number[][];
+    get model(): string | null;
+    get promptTokens(): number | null;
+    get length(): number;
+}
 
 /**
  * `client.chat.completions` — OpenAI-compatible chat completions.
  *
- * `model` is an endpoint id from `endpoints.deploy(...)` (or any model id the
- * caller's org can reach). The call is metered: a successful completion debits
- * the org's balance; a zero balance raises `InsufficientCreditsError` (402).
+ * `model` is `"auto"` — Pareta plans, routes, and verifies each request. The
+ * call is metered: a successful completion debits the org's balance; a zero
+ * balance raises `InsufficientCreditsError` (402).
  */
 
 /** An OpenAI-style chat message. `content` is a string or content-block array. */
@@ -200,7 +183,7 @@ interface ChatMessage {
 }
 /** Params for `chat.completions.create`. Extra OpenAI params pass through. */
 interface ChatCompletionCreateParams {
-    /** An endpoint id from `endpoints.deploy(...)`. */
+    /** `"auto"` — Pareta routes the request to the right model. */
     model: string;
     messages: ChatMessage[];
     stream?: boolean;
@@ -230,10 +213,9 @@ declare class Chat {
 }
 
 /**
- * `client.models` — list the deployed models (endpoints) your org can call.
- *
- * OpenAI-compatible: only deployed, url-bearing endpoints appear. Each
- * `Model.id` is an endpoint id usable as `chat.completions.create({ model })`.
+ * `client.models` — the OpenAI-compatible `/v1/models` list of model ids your
+ * org can call. Each `Model.id` is usable as
+ * `chat.completions.create({ model })`.
  */
 
 declare class Models {
@@ -243,70 +225,8 @@ declare class Models {
 }
 
 /**
- * `client.endpoints` — deploy, operate, and measure deployed endpoints.
- *
- * `deploy({ task, model })` is ergonomic: Pareta picks the GPU/serving config
- * (you never pass hardware) and `model` defaults to the task's recommended pick.
- * The backend resolver behind POST /v1/endpoints evolved with the discovery
- * work — verify the deploy body against staging when wiring real deploys.
- */
-
-/** Params for `endpoints.deploy`. Extra keys pass through to the deploy body. */
-interface DeployParams {
-    /** The task id to deploy a model for. */
-    task: string;
-    /** Model alias to deploy; defaults to "recommended" (the task's top pick). */
-    model?: string;
-    /** Optional endpoint name. */
-    name?: string;
-    /** If true, block through the deploy and resolve the live Endpoint. */
-    wait?: boolean;
-    [key: string]: unknown;
-}
-/**
- * `endpoints.metrics(id).performance()` etc. Each returns the raw metric JSON
- * (shapes vary by dimension; typed models arrive with the OpenAPI generation).
- */
-declare class EndpointMetrics {
-    private readonly client;
-    private readonly id;
-    constructor(client: Transport, id: string);
-    performance(params?: Record<string, unknown>): Promise<unknown>;
-    uptime(params?: Record<string, unknown>): Promise<unknown>;
-    cost(params?: Record<string, unknown>): Promise<unknown>;
-    quality(params?: Record<string, unknown>): Promise<unknown>;
-    activity(params?: Record<string, unknown>): Promise<unknown>;
-}
-declare class Endpoints {
-    private readonly client;
-    constructor(client: Transport);
-    /**
-     * Deploy a model for a task. `wait: false` (default) → an `AsyncIterable` of
-     * progress events (`{event, data}`); the terminal event is `complete`
-     * (`data.endpoint`) or `error`. `wait: true` → resolves the live `Endpoint`
-     * (throws `ParetaError` on a deploy `error` event).
-     */
-    deploy(params: DeployParams & {
-        wait: true;
-    }): Promise<Endpoint>;
-    deploy(params: DeployParams & {
-        wait?: false;
-    }): AsyncIterable<SSEEvent>;
-    private waitForDeploy;
-    list(): Promise<Endpoint[]>;
-    retrieve(endpointId: string): Promise<Endpoint>;
-    start(endpointId: string): Promise<unknown>;
-    stop(endpointId: string): Promise<unknown>;
-    delete(endpointId: string): Promise<void>;
-    metrics(endpointId: string): EndpointMetrics;
-}
-
-/**
- * `client.tasks` — browse the benchmark catalog + match free-text intent, and
- * the per-task discovery surface (leaderboard / recommended).
- *
- * Unlike the Python SDK (where `leaderboard`/`recommended` exist only on the
- * sync client), these are present uniformly here — JS has one client.
+ * `client.tasks` — browse the benchmark catalog and match free-text intent to
+ * a task (`match` is the discovery surface: it tells you what Pareta can do).
  */
 
 declare class Tasks {
@@ -320,10 +240,6 @@ declare class Tasks {
     match(query: string, opts?: {
         topK?: number;
     }): Promise<TaskMatch>;
-    /** Models ranked by quality/cost for a task (+ recommended + frontier baseline). */
-    leaderboard(taskId: string): Promise<Leaderboard>;
-    /** The task's recommended deployable model — what deploy(model:"recommended") resolves to. */
-    recommended(taskId: string): Promise<string | null>;
 }
 
 /**
@@ -474,6 +390,34 @@ declare class Auto {
     }): Promise<FrontierComparison>;
 }
 
+/**
+ * `pa.rerank(...)` + `pa.embeddings(...)` — the Retrieval capability lanes
+ * (the RAG stack: embeddings for recall, reranking for precision).
+ *
+ * Both are dedicated routes, not `chat.completions`:
+ *   POST /v1/rerank      {query, documents[], top_n?} -> ordered
+ *                        {index, relevance_score} rows (calibrated P(relevant))
+ *   POST /v1/embeddings  {input[], input_type?}       -> unit-normalized vectors
+ *
+ * Rerank is metered per document scored; embeddings per input token. You never
+ * pick a serving model — Pareta resolves the lane (`pareta-rerank-1` / `bge-1`),
+ * exactly as `model:"auto"` does for chat.
+ *
+ * Exposed as callable client fields (`pa.rerank(...)`) rather than resource
+ * objects — call-parity with the Python SDK, and neither lane has sub-methods.
+ */
+
+interface RerankOptions {
+    /** Truncate the response to the best N (all documents are still scored and metered). */
+    topN?: number;
+}
+type RerankFn = (query: string, documents: string[], opts?: RerankOptions) => Promise<Rerank>;
+interface EmbeddingsOptions {
+    /** `"query"` embeds a retrieval QUERY (asymmetric); default embeds documents raw. */
+    inputType?: "query" | "document";
+}
+type EmbeddingsFn = (input: string | string[], opts?: EmbeddingsOptions) => Promise<Embeddings>;
+
 /**
  * The `Pareta` client — one class (no sync/async split; JS is Promise-only).
  * A faithful port of the Python `_client.py` transport: header construction,
@@ -525,10 +469,13 @@ declare class Pareta implements Transport {
     private readonly fetchImpl;
     readonly chat: Chat;
     readonly models: Models;
-    readonly endpoints: Endpoints;
     readonly tasks: Tasks;
     readonly evals: Evals;
     readonly auto: Auto;
+    /** Document reranking (the Retrieval precision lane) — `pa.rerank(query, docs, {topN})`. */
+    readonly rerank: RerankFn;
+    /** Text embeddings (the Retrieval recall lane) — `pa.embeddings(input, {inputType})`. */
+    readonly embeddings: EmbeddingsFn;
     constructor(options?: ParetaOptions);
     /** Build from PARETA_API_KEY (+ optional PARETA_BASE_URL); explicit opts win. */
     static fromEnv(options?: ParetaOptions): Pareta;
@@ -547,12 +494,12 @@ declare class Pareta implements Transport {
     /**
      * Yield parsed SSE objects. Retries ONLY the initial connect/handshake — once
      * bytes are flowing a mid-stream drop raises (re-issuing would re-run a
-     * metered generation / re-trigger a deploy). `events:true` → `{event,data}`.
+     * metered generation). `events:true` → `{event,data}`.
      */
     stream<T = unknown>(method: string, path: string, opts?: StreamOptions): AsyncGenerator<T>;
 }
 
-declare const VERSION = "0.1.0";
+declare const VERSION = "1.1.0";
 
 /**
  * Typed error hierarchy for the Pareta SDK — a faithful port of the Python
@@ -567,7 +514,7 @@ declare const VERSION = "0.1.0";
  *   402 → InsufficientCreditsError ("organization is out of credit…")
  *   403 → PermissionDeniedError
  *   404 → NotFoundError
- *   409 → ConflictError            (seed/legacy endpoint not deployed)
+ *   409 → ConflictError            (resource conflict / not servable)
  *   422 → BadRequestError          (FastAPI validation; `detail` is an array)
  *   429 → RateLimitError
  *   503 → EndpointNotReadyError    (stopped / cold / provider down)
@@ -644,4 +591,4 @@ declare class EndpointNotReadyError extends APIStatusError {
     constructor(message: string, init: APIStatusErrorInit);
 }
 
-export { APIConnectionError, APIStatusError, APITimeoutError, AuthenticationError, BadRequestError, BaseModel, ChatCompletion, ChatCompletionChunk, type ChatCompletionCreateParams, type ChatMessage, Choice, ConflictError, type DeployParams, Endpoint, EndpointMetrics, EndpointNotReadyError, type ErrorDetail, EvalItemResult, EvalResult, EvalRun, type EvalRunCreateParams, EvalRuns, EvalSet, EvalSets, type FileInput, type FilePart, FrontierModel, type FrontierSpec, InsufficientCreditsError, Leaderboard, LeaderboardEntry, Message, Model, ModelList, NotFoundError, Pareta, ParetaError, type ParetaOptions, PermissionDeniedError, RateLimitError, type SSEEvent, Task, TaskMatch, TaskMatchCandidate, type Transport, Usage, VERSION, type ValidationErrorItem };
+export { APIConnectionError, APIStatusError, APITimeoutError, AuthenticationError, BadRequestError, BaseModel, ChatCompletion, ChatCompletionChunk, type ChatCompletionCreateParams, type ChatMessage, Choice, ConflictError, Embeddings, type EmbeddingsFn, type EmbeddingsOptions, EndpointNotReadyError, type ErrorDetail, EvalItemResult, EvalResult, EvalRun, type EvalRunCreateParams, EvalRuns, EvalSet, EvalSets, type FileInput, type FilePart, FrontierModel, type FrontierSpec, InsufficientCreditsError, Message, Model, ModelList, NotFoundError, Pareta, ParetaError, type ParetaOptions, PermissionDeniedError, RateLimitError, Rerank, type RerankFn, type RerankOptions, RerankResult, type SSEEvent, Task, TaskMatch, TaskMatchCandidate, type Transport, Usage, VERSION, type ValidationErrorItem };

package/dist/index.d.ts

@@ -11,9 +11,8 @@ interface SSEEvent {
  * typed layer.
  *
  * Alias / D3 boundary (a BACKEND contract, the SDK adds no alias logic):
- * `Endpoint.model`, `EvalResult.modelId`, leaderboard names and
- * `run.candidateModels` are per-task PUBLIC aliases — real open-weights ids
- * never cross. Frontier (vendor) ids are in the clear.
+ * `EvalResult.modelId` and `run.candidateModels` are per-task PUBLIC aliases —
+ * real open-weights ids never cross. Frontier (vendor) ids are in the clear.
  */
 type Raw = Record<string, any>;
 /** Base: holds the raw payload, dict-style escape hatches. */
@@ -63,29 +62,6 @@ declare class ModelList extends BaseModel implements Iterable<Model> {
     get length(): number;
     [Symbol.iterator](): Iterator<Model>;
 }
-/**
- * A deployed endpoint. `id` (== name) is what you pass to
- * `chat.completions.create({ model })`. `model` is the per-task public alias.
- */
-declare class Endpoint extends BaseModel {
-    get id(): string | null;
-    get name(): string | null;
-    get model(): string | null;
-    get status(): string | null;
-    get task(): string | null;
-    get url(): string | null;
-    get isLive(): boolean;
-    /** Structured-extraction endpoints (contract / SEC / ICD …): the system prompt
-     *  the benchmark used. The proxy applies it automatically when you send no
-     *  `system` message (so quality matches the leaderboard by default); returned
-     *  here so you can read it, or override it with your own `system` message.
-     *  `null` when the task has no injectable prompt (see `promptScaffold`). */
-    get recommendedSystemPrompt(): string | null;
-    /** Fixed-label classification endpoints: a copy-and-customize `system` prompt
-     *  template. NEVER auto-applied — the categories are yours, not the benchmark's
-     *  — so fill them in and send it as your `system` message. `null` otherwise. */
-    get promptScaffold(): string | null;
-}
 declare class Task extends BaseModel {
     get id(): string | null;
     get defaultScorer(): string | null;
@@ -139,29 +115,12 @@ declare class EvalResult extends BaseModel {
     /** Per-item rows (idx/score/prediction/error); empty when not persisted. */
     get perItem(): EvalItemResult[];
 }
-declare class LeaderboardEntry extends BaseModel {
-    get name(): string | null;
-    get kind(): string | null;
-    get quality(): number | null;
-    get costPerRequestMicroUsd(): number | null;
-    get contextK(): number | null;
-    get runMode(): string | null;
-}
-/** Models ranked for a task. `recommended` is the deployable pick; `frontier` the baseline. */
-declare class Leaderboard extends BaseModel {
-    get taskId(): string | null;
-    get metric(): string | null;
-    get costUnit(): string | null;
-    get recommended(): string | null;
-    get models(): LeaderboardEntry[];
-    get frontier(): LeaderboardEntry | null;
-}
 /** A vendor frontier model you can evaluate against (from the eval pool). */
 declare class FrontierModel extends BaseModel {
     get id(): string | null;
     get vendor(): string | null;
     get vision(): boolean;
-    /** Only meaningful when a task was given: it's on that task's leaderboard. */
+    /** Only meaningful when a task was given: it has benchmark results on that task. */
     get benchmarked(): boolean;
 }
 /**
@@ -183,13 +142,37 @@ declare class EvalRun extends BaseModel {
     get cost(): string;
     get results(): EvalResult[];
 }
+/** One row of `Rerank.results` — a document's position + calibrated score. */
+declare class RerankResult extends BaseModel {
+    /** Position of this document in YOUR request's documents array. */
+    get index(): number;
+    /** Calibrated P(relevant) in (0, 1) — thresholdable, not just ordinal. */
+    get relevanceScore(): number;
+}
+/** Result of `pa.rerank(...)`: `.results` ordered most-relevant-first;
+ * `.pairs` is the number of documents scored (the metered unit). */
+declare class Rerank extends BaseModel {
+    get results(): RerankResult[];
+    get model(): string | null;
+    get pairs(): number | null;
+    /** Map the ranked indices back onto the documents you sent — best first. */
+    topDocuments(documents: string[]): string[];
+}
+/** Result of `pa.embeddings(...)`: unit-normalized vectors in input order
+ * (cosine similarity is a plain dot product); `.promptTokens` is metered. */
+declare class Embeddings extends BaseModel {
+    get vectors(): number[][];
+    get model(): string | null;
+    get promptTokens(): number | null;
+    get length(): number;
+}
 
 /**
  * `client.chat.completions` — OpenAI-compatible chat completions.
  *
- * `model` is an endpoint id from `endpoints.deploy(...)` (or any model id the
- * caller's org can reach). The call is metered: a successful completion debits
- * the org's balance; a zero balance raises `InsufficientCreditsError` (402).
+ * `model` is `"auto"` — Pareta plans, routes, and verifies each request. The
+ * call is metered: a successful completion debits the org's balance; a zero
+ * balance raises `InsufficientCreditsError` (402).
  */
 
 /** An OpenAI-style chat message. `content` is a string or content-block array. */
@@ -200,7 +183,7 @@ interface ChatMessage {
 }
 /** Params for `chat.completions.create`. Extra OpenAI params pass through. */
 interface ChatCompletionCreateParams {
-    /** An endpoint id from `endpoints.deploy(...)`. */
+    /** `"auto"` — Pareta routes the request to the right model. */
     model: string;
     messages: ChatMessage[];
     stream?: boolean;
@@ -230,10 +213,9 @@ declare class Chat {
 }
 
 /**
- * `client.models` — list the deployed models (endpoints) your org can call.
- *
- * OpenAI-compatible: only deployed, url-bearing endpoints appear. Each
- * `Model.id` is an endpoint id usable as `chat.completions.create({ model })`.
+ * `client.models` — the OpenAI-compatible `/v1/models` list of model ids your
+ * org can call. Each `Model.id` is usable as
+ * `chat.completions.create({ model })`.
  */
 
 declare class Models {
@@ -243,70 +225,8 @@ declare class Models {
 }
 
 /**
- * `client.endpoints` — deploy, operate, and measure deployed endpoints.
- *
- * `deploy({ task, model })` is ergonomic: Pareta picks the GPU/serving config
- * (you never pass hardware) and `model` defaults to the task's recommended pick.
- * The backend resolver behind POST /v1/endpoints evolved with the discovery
- * work — verify the deploy body against staging when wiring real deploys.
- */
-
-/** Params for `endpoints.deploy`. Extra keys pass through to the deploy body. */
-interface DeployParams {
-    /** The task id to deploy a model for. */
-    task: string;
-    /** Model alias to deploy; defaults to "recommended" (the task's top pick). */
-    model?: string;
-    /** Optional endpoint name. */
-    name?: string;
-    /** If true, block through the deploy and resolve the live Endpoint. */
-    wait?: boolean;
-    [key: string]: unknown;
-}
-/**
- * `endpoints.metrics(id).performance()` etc. Each returns the raw metric JSON
- * (shapes vary by dimension; typed models arrive with the OpenAPI generation).
- */
-declare class EndpointMetrics {
-    private readonly client;
-    private readonly id;
-    constructor(client: Transport, id: string);
-    performance(params?: Record<string, unknown>): Promise<unknown>;
-    uptime(params?: Record<string, unknown>): Promise<unknown>;
-    cost(params?: Record<string, unknown>): Promise<unknown>;
-    quality(params?: Record<string, unknown>): Promise<unknown>;
-    activity(params?: Record<string, unknown>): Promise<unknown>;
-}
-declare class Endpoints {
-    private readonly client;
-    constructor(client: Transport);
-    /**
-     * Deploy a model for a task. `wait: false` (default) → an `AsyncIterable` of
-     * progress events (`{event, data}`); the terminal event is `complete`
-     * (`data.endpoint`) or `error`. `wait: true` → resolves the live `Endpoint`
-     * (throws `ParetaError` on a deploy `error` event).
-     */
-    deploy(params: DeployParams & {
-        wait: true;
-    }): Promise<Endpoint>;
-    deploy(params: DeployParams & {
-        wait?: false;
-    }): AsyncIterable<SSEEvent>;
-    private waitForDeploy;
-    list(): Promise<Endpoint[]>;
-    retrieve(endpointId: string): Promise<Endpoint>;
-    start(endpointId: string): Promise<unknown>;
-    stop(endpointId: string): Promise<unknown>;
-    delete(endpointId: string): Promise<void>;
-    metrics(endpointId: string): EndpointMetrics;
-}
-
-/**
- * `client.tasks` — browse the benchmark catalog + match free-text intent, and
- * the per-task discovery surface (leaderboard / recommended).
- *
- * Unlike the Python SDK (where `leaderboard`/`recommended` exist only on the
- * sync client), these are present uniformly here — JS has one client.
+ * `client.tasks` — browse the benchmark catalog and match free-text intent to
+ * a task (`match` is the discovery surface: it tells you what Pareta can do).
  */
 
 declare class Tasks {
@@ -320,10 +240,6 @@ declare class Tasks {
     match(query: string, opts?: {
         topK?: number;
     }): Promise<TaskMatch>;
-    /** Models ranked by quality/cost for a task (+ recommended + frontier baseline). */
-    leaderboard(taskId: string): Promise<Leaderboard>;
-    /** The task's recommended deployable model — what deploy(model:"recommended") resolves to. */
-    recommended(taskId: string): Promise<string | null>;
 }
 
 /**
@@ -474,6 +390,34 @@ declare class Auto {
     }): Promise<FrontierComparison>;
 }
 
+/**
+ * `pa.rerank(...)` + `pa.embeddings(...)` — the Retrieval capability lanes
+ * (the RAG stack: embeddings for recall, reranking for precision).
+ *
+ * Both are dedicated routes, not `chat.completions`:
+ *   POST /v1/rerank      {query, documents[], top_n?} -> ordered
+ *                        {index, relevance_score} rows (calibrated P(relevant))
+ *   POST /v1/embeddings  {input[], input_type?}       -> unit-normalized vectors
+ *
+ * Rerank is metered per document scored; embeddings per input token. You never
+ * pick a serving model — Pareta resolves the lane (`pareta-rerank-1` / `bge-1`),
+ * exactly as `model:"auto"` does for chat.
+ *
+ * Exposed as callable client fields (`pa.rerank(...)`) rather than resource
+ * objects — call-parity with the Python SDK, and neither lane has sub-methods.
+ */
+
+interface RerankOptions {
+    /** Truncate the response to the best N (all documents are still scored and metered). */
+    topN?: number;
+}
+type RerankFn = (query: string, documents: string[], opts?: RerankOptions) => Promise<Rerank>;
+interface EmbeddingsOptions {
+    /** `"query"` embeds a retrieval QUERY (asymmetric); default embeds documents raw. */
+    inputType?: "query" | "document";
+}
+type EmbeddingsFn = (input: string | string[], opts?: EmbeddingsOptions) => Promise<Embeddings>;
+
 /**
  * The `Pareta` client — one class (no sync/async split; JS is Promise-only).
  * A faithful port of the Python `_client.py` transport: header construction,
@@ -525,10 +469,13 @@ declare class Pareta implements Transport {
     private readonly fetchImpl;
     readonly chat: Chat;
     readonly models: Models;
-    readonly endpoints: Endpoints;
     readonly tasks: Tasks;
     readonly evals: Evals;
     readonly auto: Auto;
+    /** Document reranking (the Retrieval precision lane) — `pa.rerank(query, docs, {topN})`. */
+    readonly rerank: RerankFn;
+    /** Text embeddings (the Retrieval recall lane) — `pa.embeddings(input, {inputType})`. */
+    readonly embeddings: EmbeddingsFn;
     constructor(options?: ParetaOptions);
     /** Build from PARETA_API_KEY (+ optional PARETA_BASE_URL); explicit opts win. */
     static fromEnv(options?: ParetaOptions): Pareta;
@@ -547,12 +494,12 @@ declare class Pareta implements Transport {
     /**
      * Yield parsed SSE objects. Retries ONLY the initial connect/handshake — once
      * bytes are flowing a mid-stream drop raises (re-issuing would re-run a
-     * metered generation / re-trigger a deploy). `events:true` → `{event,data}`.
+     * metered generation). `events:true` → `{event,data}`.
      */
     stream<T = unknown>(method: string, path: string, opts?: StreamOptions): AsyncGenerator<T>;
 }
 
-declare const VERSION = "0.1.0";
+declare const VERSION = "1.1.0";
 
 /**
  * Typed error hierarchy for the Pareta SDK — a faithful port of the Python
@@ -567,7 +514,7 @@ declare const VERSION = "0.1.0";
  *   402 → InsufficientCreditsError ("organization is out of credit…")
  *   403 → PermissionDeniedError
  *   404 → NotFoundError
- *   409 → ConflictError            (seed/legacy endpoint not deployed)
+ *   409 → ConflictError            (resource conflict / not servable)
  *   422 → BadRequestError          (FastAPI validation; `detail` is an array)
  *   429 → RateLimitError
  *   503 → EndpointNotReadyError    (stopped / cold / provider down)
@@ -644,4 +591,4 @@ declare class EndpointNotReadyError extends APIStatusError {
     constructor(message: string, init: APIStatusErrorInit);
 }
 
-export { APIConnectionError, APIStatusError, APITimeoutError, AuthenticationError, BadRequestError, BaseModel, ChatCompletion, ChatCompletionChunk, type ChatCompletionCreateParams, type ChatMessage, Choice, ConflictError, type DeployParams, Endpoint, EndpointMetrics, EndpointNotReadyError, type ErrorDetail, EvalItemResult, EvalResult, EvalRun, type EvalRunCreateParams, EvalRuns, EvalSet, EvalSets, type FileInput, type FilePart, FrontierModel, type FrontierSpec, InsufficientCreditsError, Leaderboard, LeaderboardEntry, Message, Model, ModelList, NotFoundError, Pareta, ParetaError, type ParetaOptions, PermissionDeniedError, RateLimitError, type SSEEvent, Task, TaskMatch, TaskMatchCandidate, type Transport, Usage, VERSION, type ValidationErrorItem };
+export { APIConnectionError, APIStatusError, APITimeoutError, AuthenticationError, BadRequestError, BaseModel, ChatCompletion, ChatCompletionChunk, type ChatCompletionCreateParams, type ChatMessage, Choice, ConflictError, Embeddings, type EmbeddingsFn, type EmbeddingsOptions, EndpointNotReadyError, type ErrorDetail, EvalItemResult, EvalResult, EvalRun, type EvalRunCreateParams, EvalRuns, EvalSet, EvalSets, type FileInput, type FilePart, FrontierModel, type FrontierSpec, InsufficientCreditsError, Message, Model, ModelList, NotFoundError, Pareta, ParetaError, type ParetaOptions, PermissionDeniedError, RateLimitError, Rerank, type RerankFn, type RerankOptions, RerankResult, type SSEEvent, Task, TaskMatch, TaskMatchCandidate, type Transport, Usage, VERSION, type ValidationErrorItem };