@promptev/context-engine 0.0.1 → 0.0.2

package/dist/index.d.cts

@@ -1,12 +1,12 @@
 import { Pool } from 'pg';
-import { C as ContextEngineConfig, L as LLMConfig, E as ExtractionConfig, R as RerankerConfig } from './config-CNnASw5X.cjs';
-export { a as ContextEngineConfigInit, b as EmbeddingConfig, F as FusionConfig, G as GraphConfig, S as StorageConfig } from './config-CNnASw5X.cjs';
+import { C as ContextEngineConfig, L as LLMConfig, E as ExtractionConfig, R as RerankerConfig } from './config-C5RZ00W6.cjs';
+export { a as ContextEngineConfigInit, b as EmbeddingConfig, F as FusionConfig, G as GraphConfig, S as StorageConfig } from './config-C5RZ00W6.cjs';
 import { H as Hooks, R as RedactionPolicy, U as UsageEvent, P as ProgressEvent, I as IngestReport } from './redaction-BqD_DEUQ.cjs';
 export { D as DocumentReport, a as RedactionRule, b as RedactionRuleInit, c as applyRedaction, e as emitError, d as emitProgress, f as emitToolCall, g as emitUsage, h as graphUnits, u as unitsForFile } from './redaction-BqD_DEUQ.cjs';
-import { E as Embedder, S as StorageBackend, F as FetchImpl } from './storage-DU1JRno5.cjs';
-export { C as ChunkRow, a as EmbedKind, P as PostgresBackend, b as SearchScope, c as buildEmbedder } from './storage-DU1JRno5.cjs';
-import { T as ToolEngine, C as CanonicalTool, b as ToolHttpClient, a as ToolConfig } from './governance-D8g6Wyvb.cjs';
-export { c as ToolKind, d as configSchema } from './governance-D8g6Wyvb.cjs';
+import { E as Embedder, S as StorageBackend, F as FetchImpl } from './storage-Dvpq2xAC.cjs';
+export { C as ChunkRow, a as EmbedKind, P as PostgresBackend, b as SearchScope, c as buildEmbedder } from './storage-Dvpq2xAC.cjs';
+import { T as ToolEngine, C as CanonicalTool, b as ToolHttpClient, a as ToolConfig } from './governance-P9pRb4Ol.cjs';
+export { c as ToolKind, d as configSchema } from './governance-P9pRb4Ol.cjs';
 export { createMcpApp } from './mcp.cjs';
 import 'zod';
 import 'node:http';
@@ -129,6 +129,13 @@ declare function queryStructured(question: string, opts: {
  * call, no sandbox run — when the flag is `false`. That flag, not the
  * isolate, is the actual security boundary.
  *
+ * Only the document half lives here: fetch the in-scope tabular documents
+ * and parse their stored text into `{sheet: rows[]}`. Everything after that
+ * — the guards, both redaction surfaces below, the prompt → code → sandbox
+ * path and the swept result — is `computeOverFrames`, the seam a host calls
+ * when it already holds the frames and has no document to point at; this
+ * function builds its frames from documents and delegates.
+ *
  * `config.redaction` is applied TWICE:
  * 1. Each document's raw `.text` is masked BEFORE it is parsed into a
  *    table, so the data the LLM-authored code runs against is built from
@@ -146,6 +153,54 @@ declare function compute(instruction: string, opts: {
     modelCfg?: LLMConfig | null;
     timeout?: number;
 }): Promise<Record<string, unknown>>;
+/** What `compute()` builds and `computeOverFrames` consumes: sheet name → rows. */
+type ComputeFrames = Record<string, Record<string, unknown>[]>;
+/** A source document behind a frame, as `compute()` reports it in `documentsUsed`. */
+interface ComputeDocument {
+    id: string;
+    name?: string | null;
+    sourceId?: string | null;
+}
+/**
+ * Compute an answer to `instruction` over caller-supplied tables.
+ *
+ * The frame-level seam under `compute()`: everything `compute()` does AFTER
+ * it has turned its documents into `{sheet: rows[]}` lives here, so a host
+ * that already holds the frames — an uploaded workbook, a connector's
+ * sheet, a query result — can run the same prompt → code → sandbox path
+ * without first ingesting a document to point at.
+ *
+ * `frames` keys are the sheet names the caller chose; the LLM sees them and
+ * each row's columns exactly as it sees a parsed document's.
+ *
+ * The guards are `compute()`'s, in the same order and all BEFORE any LLM
+ * call: `config.enableCodeExecution` off → `EngineActionError`; no LLM
+ * (`modelCfg` or `config.llm`) → `Error`; blank instruction or empty
+ * `frames` → `EngineActionError`; `timeout` clamped to 1..300 seconds.
+ *
+ * `config.redaction` is applied at the same two surfaces as `compute()`:
+ *
+ * 1. Every string cell and every column name is masked BEFORE the prompt is
+ *    built (`maskFrames`), so neither the schema summary the LLM reads nor
+ *    the rows its code runs against carry a raw value. Same intended
+ *    trade-off as `compute()`'s point 1 — a masked cell can change a
+ *    computed result, and that is correct.
+ * 2. The returned object is swept whole through `redactValueRecursive`,
+ *    `code` included — `compute()`'s point 2, deliberately blunt.
+ *
+ * `hooks`, `principals` and `documents` are how `compute()` threads its own
+ * context through; a host calling the seam directly normally leaves them at
+ * their defaults (no error hook, trusted-internal `unless` evaluation, no
+ * source documents — `documentsUsed` comes back empty).
+ */
+declare function computeOverFrames(frames: ComputeFrames, instruction: string, opts: {
+    config: ContextEngineConfig;
+    modelCfg?: LLMConfig | null;
+    timeout?: number;
+    hooks?: Hooks | null;
+    principals?: string[] | null;
+    documents?: ComputeDocument[] | null;
+}): Promise<Record<string, unknown>>;
 
 /**
  * AES-256-GCM crypto seam. Wire format is a compatibility promise with the
@@ -811,4 +866,4 @@ declare function functionTool(fn: (...args: never[]) => unknown): CanonicalTool;
 /** Bumped by CI on every main merge; 0.0.0 = pre-first-release. */
 declare const __version__ = "0.0.0";
 
-export { ApprovalExpired, ApprovalNotPending, type ApprovalRecord, CeleryRunner, CodeExecutionError, CodeExecutionTimeout, ContextEngine, ContextEngineConfig, DEFAULT_LEG_WEIGHT, DocumentNotFoundError, EXTRACTION_VERSION, Embedder, EngineActionError, ExtraMissingError, Extracted, ExtractionConfig, type ExtractionResult, GraphLegUnavailable, type Hit, Hooks, InProcessRunner, IngestReport, LLMClient, LLMConfig, type Principals, ProgressEvent, RedactionPolicy, RerankerConfig, type SearchResult, StorageBackend, TRUSTED, type TaskRunner, type TaskStatus, ToolConfig, type Trusted, UNSET, type Unset, UsageEvent, __version__, buildLlmClient, callLlm, compute, decryptDict, encryptDict, extract, extractStructuredData, functionTool, getDocumentText, getSecretKey, listDocuments, queryStructured, redactHits, rerank, resolveApproval, resolveFields, resolvePrincipals, rrfFuse, runMigrate, runSearch, shouldRequireApproval, upsertRegistry };
+export { ApprovalExpired, ApprovalNotPending, type ApprovalRecord, CeleryRunner, CodeExecutionError, CodeExecutionTimeout, type ComputeDocument, type ComputeFrames, ContextEngine, ContextEngineConfig, DEFAULT_LEG_WEIGHT, DocumentNotFoundError, EXTRACTION_VERSION, Embedder, EngineActionError, ExtraMissingError, Extracted, ExtractionConfig, type ExtractionResult, GraphLegUnavailable, type Hit, Hooks, InProcessRunner, IngestReport, LLMClient, LLMConfig, type Principals, ProgressEvent, RedactionPolicy, RerankerConfig, type SearchResult, StorageBackend, TRUSTED, type TaskRunner, type TaskStatus, ToolConfig, type Trusted, UNSET, type Unset, UsageEvent, __version__, buildLlmClient, callLlm, compute, computeOverFrames, decryptDict, encryptDict, extract, extractStructuredData, functionTool, getDocumentText, getSecretKey, listDocuments, queryStructured, redactHits, rerank, resolveApproval, resolveFields, resolvePrincipals, rrfFuse, runMigrate, runSearch, shouldRequireApproval, upsertRegistry };

package/dist/index.d.ts

@@ -1,12 +1,12 @@
 import { Pool } from 'pg';
-import { C as ContextEngineConfig, L as LLMConfig, E as ExtractionConfig, R as RerankerConfig } from './config-CdlSkKgV.js';
-export { a as ContextEngineConfigInit, b as EmbeddingConfig, F as FusionConfig, G as GraphConfig, S as StorageConfig } from './config-CdlSkKgV.js';
+import { C as ContextEngineConfig, L as LLMConfig, E as ExtractionConfig, R as RerankerConfig } from './config-BODDdXJ7.js';
+export { a as ContextEngineConfigInit, b as EmbeddingConfig, F as FusionConfig, G as GraphConfig, S as StorageConfig } from './config-BODDdXJ7.js';
 import { H as Hooks, R as RedactionPolicy, U as UsageEvent, P as ProgressEvent, I as IngestReport } from './redaction-BqD_DEUQ.js';
 export { D as DocumentReport, a as RedactionRule, b as RedactionRuleInit, c as applyRedaction, e as emitError, d as emitProgress, f as emitToolCall, g as emitUsage, h as graphUnits, u as unitsForFile } from './redaction-BqD_DEUQ.js';
-import { E as Embedder, S as StorageBackend, F as FetchImpl } from './storage-Dvt2ZxsV.js';
-export { C as ChunkRow, a as EmbedKind, P as PostgresBackend, b as SearchScope, c as buildEmbedder } from './storage-Dvt2ZxsV.js';
-import { T as ToolEngine, C as CanonicalTool, b as ToolHttpClient, a as ToolConfig } from './governance-XFVgtEdV.js';
-export { c as ToolKind, d as configSchema } from './governance-XFVgtEdV.js';
+import { E as Embedder, S as StorageBackend, F as FetchImpl } from './storage-CJrKgJeJ.js';
+export { C as ChunkRow, a as EmbedKind, P as PostgresBackend, b as SearchScope, c as buildEmbedder } from './storage-CJrKgJeJ.js';
+import { T as ToolEngine, C as CanonicalTool, b as ToolHttpClient, a as ToolConfig } from './governance-BLPK7NMe.js';
+export { c as ToolKind, d as configSchema } from './governance-BLPK7NMe.js';
 export { createMcpApp } from './mcp.js';
 import 'zod';
 import 'node:http';
@@ -129,6 +129,13 @@ declare function queryStructured(question: string, opts: {
  * call, no sandbox run — when the flag is `false`. That flag, not the
  * isolate, is the actual security boundary.
  *
+ * Only the document half lives here: fetch the in-scope tabular documents
+ * and parse their stored text into `{sheet: rows[]}`. Everything after that
+ * — the guards, both redaction surfaces below, the prompt → code → sandbox
+ * path and the swept result — is `computeOverFrames`, the seam a host calls
+ * when it already holds the frames and has no document to point at; this
+ * function builds its frames from documents and delegates.
+ *
  * `config.redaction` is applied TWICE:
  * 1. Each document's raw `.text` is masked BEFORE it is parsed into a
  *    table, so the data the LLM-authored code runs against is built from
@@ -146,6 +153,54 @@ declare function compute(instruction: string, opts: {
     modelCfg?: LLMConfig | null;
     timeout?: number;
 }): Promise<Record<string, unknown>>;
+/** What `compute()` builds and `computeOverFrames` consumes: sheet name → rows. */
+type ComputeFrames = Record<string, Record<string, unknown>[]>;
+/** A source document behind a frame, as `compute()` reports it in `documentsUsed`. */
+interface ComputeDocument {
+    id: string;
+    name?: string | null;
+    sourceId?: string | null;
+}
+/**
+ * Compute an answer to `instruction` over caller-supplied tables.
+ *
+ * The frame-level seam under `compute()`: everything `compute()` does AFTER
+ * it has turned its documents into `{sheet: rows[]}` lives here, so a host
+ * that already holds the frames — an uploaded workbook, a connector's
+ * sheet, a query result — can run the same prompt → code → sandbox path
+ * without first ingesting a document to point at.
+ *
+ * `frames` keys are the sheet names the caller chose; the LLM sees them and
+ * each row's columns exactly as it sees a parsed document's.
+ *
+ * The guards are `compute()`'s, in the same order and all BEFORE any LLM
+ * call: `config.enableCodeExecution` off → `EngineActionError`; no LLM
+ * (`modelCfg` or `config.llm`) → `Error`; blank instruction or empty
+ * `frames` → `EngineActionError`; `timeout` clamped to 1..300 seconds.
+ *
+ * `config.redaction` is applied at the same two surfaces as `compute()`:
+ *
+ * 1. Every string cell and every column name is masked BEFORE the prompt is
+ *    built (`maskFrames`), so neither the schema summary the LLM reads nor
+ *    the rows its code runs against carry a raw value. Same intended
+ *    trade-off as `compute()`'s point 1 — a masked cell can change a
+ *    computed result, and that is correct.
+ * 2. The returned object is swept whole through `redactValueRecursive`,
+ *    `code` included — `compute()`'s point 2, deliberately blunt.
+ *
+ * `hooks`, `principals` and `documents` are how `compute()` threads its own
+ * context through; a host calling the seam directly normally leaves them at
+ * their defaults (no error hook, trusted-internal `unless` evaluation, no
+ * source documents — `documentsUsed` comes back empty).
+ */
+declare function computeOverFrames(frames: ComputeFrames, instruction: string, opts: {
+    config: ContextEngineConfig;
+    modelCfg?: LLMConfig | null;
+    timeout?: number;
+    hooks?: Hooks | null;
+    principals?: string[] | null;
+    documents?: ComputeDocument[] | null;
+}): Promise<Record<string, unknown>>;
 
 /**
  * AES-256-GCM crypto seam. Wire format is a compatibility promise with the
@@ -811,4 +866,4 @@ declare function functionTool(fn: (...args: never[]) => unknown): CanonicalTool;
 /** Bumped by CI on every main merge; 0.0.0 = pre-first-release. */
 declare const __version__ = "0.0.0";
 
-export { ApprovalExpired, ApprovalNotPending, type ApprovalRecord, CeleryRunner, CodeExecutionError, CodeExecutionTimeout, ContextEngine, ContextEngineConfig, DEFAULT_LEG_WEIGHT, DocumentNotFoundError, EXTRACTION_VERSION, Embedder, EngineActionError, ExtraMissingError, Extracted, ExtractionConfig, type ExtractionResult, GraphLegUnavailable, type Hit, Hooks, InProcessRunner, IngestReport, LLMClient, LLMConfig, type Principals, ProgressEvent, RedactionPolicy, RerankerConfig, type SearchResult, StorageBackend, TRUSTED, type TaskRunner, type TaskStatus, ToolConfig, type Trusted, UNSET, type Unset, UsageEvent, __version__, buildLlmClient, callLlm, compute, decryptDict, encryptDict, extract, extractStructuredData, functionTool, getDocumentText, getSecretKey, listDocuments, queryStructured, redactHits, rerank, resolveApproval, resolveFields, resolvePrincipals, rrfFuse, runMigrate, runSearch, shouldRequireApproval, upsertRegistry };
+export { ApprovalExpired, ApprovalNotPending, type ApprovalRecord, CeleryRunner, CodeExecutionError, CodeExecutionTimeout, type ComputeDocument, type ComputeFrames, ContextEngine, ContextEngineConfig, DEFAULT_LEG_WEIGHT, DocumentNotFoundError, EXTRACTION_VERSION, Embedder, EngineActionError, ExtraMissingError, Extracted, ExtractionConfig, type ExtractionResult, GraphLegUnavailable, type Hit, Hooks, InProcessRunner, IngestReport, LLMClient, LLMConfig, type Principals, ProgressEvent, RedactionPolicy, RerankerConfig, type SearchResult, StorageBackend, TRUSTED, type TaskRunner, type TaskStatus, ToolConfig, type Trusted, UNSET, type Unset, UsageEvent, __version__, buildLlmClient, callLlm, compute, computeOverFrames, decryptDict, encryptDict, extract, extractStructuredData, functionTool, getDocumentText, getSecretKey, listDocuments, queryStructured, redactHits, rerank, resolveApproval, resolveFields, resolvePrincipals, rrfFuse, runMigrate, runSearch, shouldRequireApproval, upsertRegistry };

package/dist/index.js

@@ -803,6 +803,47 @@ var init_hooks = __esm({
   }
 });
 
+// src/providers/google.ts
+async function buildGenaiClient(cfg2, timeoutMs, purpose) {
+  const specifier = "@google/genai";
+  let mod;
+  try {
+    mod = await import(specifier);
+  } catch {
+    throw new ExtraMissingError("gemini", specifier, purpose);
+  }
+  const Ctor = mod.GoogleGenAI ?? mod.Client;
+  if (!Ctor) {
+    throw new ExtraMissingError("gemini", specifier, purpose);
+  }
+  const opts = { httpOptions: { timeout: timeoutMs } };
+  if (cfg2.provider === "vertex_ai") {
+    opts.vertexai = true;
+    if (cfg2.project || cfg2.location) {
+      if (cfg2.project) opts.project = cfg2.project;
+      if (cfg2.location) opts.location = cfg2.location;
+    } else if (cfg2.apiKey) {
+      opts.apiKey = cfg2.apiKey;
+    }
+  } else {
+    opts.apiKey = cfg2.apiKey ?? null;
+  }
+  try {
+    return new Ctor(opts);
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err);
+    if (!message.includes("Authentication is not set up")) throw err;
+    throw new Error(
+      `${message} Set \`project\` on the provider config, or export GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION. Credentials themselves come from Application Default Credentials.`
+    );
+  }
+}
+var init_google = __esm({
+  "src/providers/google.ts"() {
+    init_errors();
+  }
+});
+
 // src/providers/llm.ts
 var llm_exports = {};
 __export(llm_exports, {
@@ -838,20 +879,6 @@ function buildOpenAIChatClient(cfg2) {
     maxRetries: 0
   });
 }
-async function loadGeminiChatClient(apiKey) {
-  const specifier = "@google/genai";
-  let mod;
-  try {
-    mod = await import(specifier);
-  } catch {
-    throw new ExtraMissingError("gemini", specifier, "gemini llm");
-  }
-  const Ctor = mod.GoogleGenAI ?? mod.Client;
-  if (!Ctor) {
-    throw new ExtraMissingError("gemini", specifier, "gemini llm");
-  }
-  return new Ctor({ apiKey: apiKey ?? null, httpOptions: { timeout: GEMINI_CALL_TIMEOUT_MS } });
-}
 async function loadBedrockSdk() {
   const specifier = "@aws-sdk/client-bedrock-runtime";
   try {
@@ -870,7 +897,7 @@ function buildLlmClient(cfg2, opts) {
   if (OPENAI_FAMILY.has(cfg2.provider)) {
     return new LLMClient(cfg2, { client: buildOpenAIChatClient(cfg2) });
   }
-  if (cfg2.provider === "gemini" || cfg2.provider === "bedrock") {
+  if (GOOGLE_FAMILY.has(cfg2.provider) || cfg2.provider === "bedrock") {
     return new LLMClient(cfg2);
   }
   throw new Error(`unknown llm provider: ${JSON.stringify(cfg2.provider)}`);
@@ -887,16 +914,18 @@ async function callLlm(cfg2, opts) {
     await owned.aclose();
   }
 }
-var CALL_TIMEOUT_MS, TIMEOUT_MS, GEMINI_CALL_TIMEOUT_MS, ANTHROPIC_VERSION, ANTHROPIC_MAX_TOKENS, OPENAI_FAMILY, LLMClient;
+var CALL_TIMEOUT_MS, TIMEOUT_MS, GEMINI_CALL_TIMEOUT_MS, ANTHROPIC_VERSION, ANTHROPIC_MAX_TOKENS, OPENAI_FAMILY, GOOGLE_FAMILY, LLMClient;
 var init_llm = __esm({
   "src/providers/llm.ts"() {
     init_errors();
+    init_google();
     CALL_TIMEOUT_MS = 24e4;
     TIMEOUT_MS = CALL_TIMEOUT_MS;
     GEMINI_CALL_TIMEOUT_MS = CALL_TIMEOUT_MS;
     ANTHROPIC_VERSION = "2023-06-01";
     ANTHROPIC_MAX_TOKENS = 4096;
     OPENAI_FAMILY = /* @__PURE__ */ new Set(["openai", "azure_openai", "custom"]);
+    GOOGLE_FAMILY = /* @__PURE__ */ new Set(["gemini", "vertex_ai"]);
     LLMClient = class {
       cfg;
       provider;
@@ -936,7 +965,7 @@ var init_llm = __esm({
         if (OPENAI_FAMILY.has(this.provider)) {
           return this.callOpenAI(system, user, jsonMode, images, maxTokens, temperature);
         }
-        if (this.provider === "gemini") {
+        if (GOOGLE_FAMILY.has(this.provider)) {
           return this.callGemini(system, user, jsonMode, images, maxTokens, thinkingBudget, temperature);
         }
         if (this.provider === "bedrock") {
@@ -1019,7 +1048,11 @@ Respond with valid JSON only.`;
       }
       async callGemini(system, user, jsonMode, images, maxTokens, thinkingBudget = null, temperature = null) {
         if (!this.genaiClient) {
-          this.genaiClient = await loadGeminiChatClient(this.cfg.apiKey);
+          this.genaiClient = await buildGenaiClient(
+            this.cfg,
+            GEMINI_CALL_TIMEOUT_MS,
+            "gemini llm"
+          );
         }
         const parts = [];
         for (const img of images ?? []) {
@@ -1898,17 +1931,30 @@ var init_config = __esm({
   "src/config.ts"() {
     init_redaction();
     embeddingSchema = z.object({
-      provider: z.enum(["openai", "azure_openai", "gemini", "voyage", "cohere", "custom"]),
+      provider: z.enum(["openai", "azure_openai", "gemini", "vertex_ai", "voyage", "cohere", "custom"]),
       model: z.string(),
       dim: z.number().int().positive().nullable().optional().default(null),
       apiKey: z.string().nullable().optional().default(null),
-      baseUrl: z.string().nullable().optional().default(null)
+      baseUrl: z.string().nullable().optional().default(null),
+      // `vertex_ai` only. Left optional on purpose: the Google SDK resolves both
+      // from GOOGLE_CLOUD_PROJECT / GOOGLE_CLOUD_LOCATION, which is how a GCP
+      // deployment is already wired, and requiring them here would break it.
+      //
+      // No `.default(null)`, unlike the fields above — these types come from
+      // `z.infer`, so a default would make them REQUIRED on the output type and
+      // break every hand-written `EmbeddingConfig` literal already compiled
+      // against 0.0.1. Optional keeps the addition additive.
+      project: z.string().nullable().optional(),
+      location: z.string().nullable().optional()
     });
     llmSchema = z.object({
-      provider: z.enum(["anthropic", "openai", "azure_openai", "gemini", "bedrock", "custom"]),
+      provider: z.enum(["anthropic", "openai", "azure_openai", "gemini", "vertex_ai", "bedrock", "custom"]),
       model: z.string(),
       apiKey: z.string().nullable().optional().default(null),
-      baseUrl: z.string().nullable().optional().default(null)
+      baseUrl: z.string().nullable().optional().default(null),
+      // `vertex_ai` only — see the note on embeddingSchema.
+      project: z.string().nullable().optional(),
+      location: z.string().nullable().optional()
     });
     graphSchema = z.object({
       enabled: z.boolean().default(false),
@@ -7567,20 +7613,8 @@ function parseSpreadsheetText(dfd, text) {
   return out;
 }
 async function compute(instruction, opts) {
-  if (!opts.config.enableCodeExecution) {
-    throw new EngineActionError(
-      "compute() executes generated code and is disabled by default; set enableCodeExecution=true only in a deployment with out-of-process/container isolation."
-    );
-  }
-  const llmCfg = opts.modelCfg ?? opts.config.llm;
-  if (llmCfg == null) {
-    throw new Error("compute() requires an LLM: pass modelCfg= or configure ContextEngineConfig.llm");
-  }
-  if (!instruction?.trim()) {
-    throw new EngineActionError("instruction must not be empty");
-  }
+  checkComputePreconditions(opts.config, opts.modelCfg, instruction);
   const dfd = await requireDanfo();
-  const timeout = Math.max(1, Math.min(Math.trunc(opts.timeout || DEFAULT_COMPUTE_TIMEOUT), 300));
   const principals = opts.principals ?? null;
   const params = [];
   let where = scopeSql(opts.sourceIds ?? null, principals, params);
@@ -7650,6 +7684,69 @@ async function compute(instruction, opts) {
   if (!Object.keys(dfs).length) {
     throw new EngineActionError("in-scope documents did not parse into any usable dataframe");
   }
+  return computeOverFrames(dfs, instruction, {
+    config: opts.config,
+    modelCfg: opts.modelCfg,
+    timeout: opts.timeout,
+    hooks: opts.hooks,
+    principals,
+    documents
+  });
+}
+function checkComputePreconditions(config, modelCfg, instruction) {
+  if (!config.enableCodeExecution) {
+    throw new EngineActionError(
+      "compute() executes generated code and is disabled by default; set enableCodeExecution=true only in a deployment with out-of-process/container isolation."
+    );
+  }
+  const llmCfg = modelCfg ?? config.llm;
+  if (llmCfg == null) {
+    throw new Error("compute() requires an LLM: pass modelCfg= or configure ContextEngineConfig.llm");
+  }
+  if (!instruction?.trim()) {
+    throw new EngineActionError("instruction must not be empty");
+  }
+  return llmCfg;
+}
+function maskFrames(frames, policy, opts) {
+  if (policy == null || policy.isEmpty()) return frames;
+  const mask = (text) => redactValueRecursive(text, policy, opts);
+  const masked = {};
+  for (const [name, rows] of Object.entries(frames)) {
+    const labels = /* @__PURE__ */ new Map();
+    const taken = /* @__PURE__ */ new Set();
+    for (const row of rows) {
+      for (const column of Object.keys(row)) {
+        if (labels.has(column)) continue;
+        const base = mask(column);
+        let label = base;
+        for (let n = 2; taken.has(label); n++) label = `${base}_${n}`;
+        taken.add(label);
+        labels.set(column, label);
+      }
+    }
+    masked[name] = rows.map(
+      (row) => Object.fromEntries(
+        Object.entries(row).map(([column, value]) => [
+          labels.get(column) ?? column,
+          typeof value === "string" ? mask(value) : value
+        ])
+      )
+    );
+  }
+  return masked;
+}
+async function computeOverFrames(frames, instruction, opts) {
+  const llmCfg = checkComputePreconditions(opts.config, opts.modelCfg, instruction);
+  if (!frames || !Object.keys(frames).length) {
+    throw new EngineActionError("no tabular data to compute over");
+  }
+  const hooks = opts.hooks ?? {};
+  const principals = opts.principals ?? null;
+  const documents = [...opts.documents ?? []];
+  const timeout = Math.max(1, Math.min(Math.trunc(opts.timeout || DEFAULT_COMPUTE_TIMEOUT), 300));
+  const redactOpts = { principals, secretKey: opts.config.secretKey, hooks };
+  const dfs = maskFrames(frames, opts.config.redaction, redactOpts);
   const schemaLines = Object.entries(dfs).map(
     ([name, table]) => `- ${name}: columns=${JSON.stringify(Object.keys(table[0] ?? {}))}, rows=${table.length}`
   );
@@ -7666,7 +7763,7 @@ ${schemaLines.join("\n")}`;
       jsonMode: false
     });
   } catch (exc) {
-    emitError(opts.hooks, exc, { stage: "compute_codegen", instruction: instruction.slice(0, 200) });
+    emitError(hooks, exc, { stage: "compute_codegen", instruction: instruction.slice(0, 200) });
     throw exc;
   }
   const code = stripCodeFences(rawCode);
@@ -7677,12 +7774,12 @@ ${schemaLines.join("\n")}`;
       maskedCode = redactValueRecursive(code.slice(0, 500), opts.config.redaction, {
         principals,
         secretKey: opts.config.secretKey,
-        hooks: opts.hooks
+        hooks
       });
     } catch {
       maskedCode = "<redaction failed: code omitted>";
     }
-    emitError(opts.hooks, new Error(execResult.error || "compute execution failed"), {
+    emitError(hooks, new Error(execResult.error || "compute execution failed"), {
       stage: "compute_exec",
       code: maskedCode
     });
@@ -7703,7 +7800,7 @@ ${schemaLines.join("\n")}`;
   return redactValueRecursive(result, opts.config.redaction, {
     principals,
     secretKey: opts.config.secretKey,
-    hooks: opts.hooks
+    hooks
   });
 }
 
@@ -7818,10 +7915,11 @@ init_errors();
 init_hooks();
 
 // src/providers/embeddings.ts
-init_errors();
 init_text();
+init_google();
 var TIMEOUT_MS3 = 3e4;
 var OPENAI_FAMILY2 = /* @__PURE__ */ new Set(["openai", "azure_openai", "custom"]);
+var GOOGLE_FAMILY2 = /* @__PURE__ */ new Set(["gemini", "vertex_ai"]);
 function estimatedTokens(texts) {
   return texts.reduce((n, t) => n + tokenCount(t), 0);
 }
@@ -7893,7 +7991,7 @@ var Embedder = class {
   /** Provider dispatch. Overridable per-instance (tests stub this). */
   async rawEmbed(texts, kind = "document") {
     if (OPENAI_FAMILY2.has(this.provider)) return this.embedOpenAI(texts);
-    if (this.provider === "gemini") return this.embedGemini(texts);
+    if (GOOGLE_FAMILY2.has(this.provider)) return this.embedGemini(texts);
     if (this.provider === "voyage") return this.embedVoyage(texts, kind);
     if (this.provider === "cohere") return this.embedCohere(texts, kind);
     throw new Error(`unknown embedding provider: ${JSON.stringify(this.provider)}`);
@@ -7909,7 +8007,7 @@ var Embedder = class {
   }
   async embedGemini(texts) {
     if (!this.genaiClient) {
-      this.genaiClient = await loadGeminiEmbedClient(this.cfg.apiKey);
+      this.genaiClient = await buildGenaiClient(this.cfg, TIMEOUT_MS3, "gemini embeddings");
     }
     const resp = await this.genaiClient.models.embedContent({
       model: this.model,
@@ -7958,20 +8056,6 @@ var Embedder = class {
     return this.fetchImpl;
   }
 };
-async function loadGeminiEmbedClient(apiKey) {
-  const specifier = "@google/genai";
-  let mod;
-  try {
-    mod = await import(specifier);
-  } catch {
-    throw new ExtraMissingError("gemini", specifier, "gemini embeddings");
-  }
-  const Ctor = mod.GoogleGenAI ?? mod.Client;
-  if (!Ctor) {
-    throw new ExtraMissingError("gemini", specifier, "gemini embeddings");
-  }
-  return new Ctor({ apiKey: apiKey ?? null, httpOptions: { timeout: TIMEOUT_MS3 } });
-}
 function buildEmbedder(cfg2, opts) {
   if (opts?.client || opts?.fetch || opts?.fetchImpl) {
     return new Embedder(cfg2, opts);
@@ -7982,7 +8066,7 @@ function buildEmbedder(cfg2, opts) {
   if (cfg2.provider === "voyage" || cfg2.provider === "cohere") {
     return new Embedder(cfg2, { fetch: globalThis.fetch });
   }
-  if (cfg2.provider === "gemini") {
+  if (GOOGLE_FAMILY2.has(cfg2.provider)) {
     return new Embedder(cfg2);
   }
   throw new Error(`unknown embedding provider: ${JSON.stringify(cfg2.provider)}`);
@@ -12949,6 +13033,6 @@ var CeleryRunner = class {
 // src/index.ts
 init_usage();
 
-export { ApprovalExpired, ApprovalNotPending, CeleryRunner, CodeExecutionError, CodeExecutionTimeout, ContextEngine, ContextEngineConfig, DEFAULT_LEG_WEIGHT, DocumentNotFoundError, EXTRACTION_VERSION, Embedder, EngineActionError, ExtraMissingError, Extracted, GraphLegUnavailable, InProcessRunner, LLMClient, PostgresBackend, RedactionPolicy, RedactionRule, TRUSTED, ToolConfig, UNSET, __version__, applyRedaction, buildEmbedder, buildLlmClient, callLlm, compute, configSchema, createMcpApp, decryptDict, emitError, emitProgress, emitToolCall, emitUsage, encryptDict, extract2 as extract, extractStructuredData, functionTool, getDocumentText, getSecretKey, graphUnits, listDocuments, queryStructured, redactHits, rerank, resolveApproval, resolveFields, resolvePrincipals, rrfFuse, runMigrate, runSearch, shouldRequireApproval, unitsForFile, upsertRegistry };
+export { ApprovalExpired, ApprovalNotPending, CeleryRunner, CodeExecutionError, CodeExecutionTimeout, ContextEngine, ContextEngineConfig, DEFAULT_LEG_WEIGHT, DocumentNotFoundError, EXTRACTION_VERSION, Embedder, EngineActionError, ExtraMissingError, Extracted, GraphLegUnavailable, InProcessRunner, LLMClient, PostgresBackend, RedactionPolicy, RedactionRule, TRUSTED, ToolConfig, UNSET, __version__, applyRedaction, buildEmbedder, buildLlmClient, callLlm, compute, computeOverFrames, configSchema, createMcpApp, decryptDict, emitError, emitProgress, emitToolCall, emitUsage, encryptDict, extract2 as extract, extractStructuredData, functionTool, getDocumentText, getSecretKey, graphUnits, listDocuments, queryStructured, redactHits, rerank, resolveApproval, resolveFields, resolvePrincipals, rrfFuse, runMigrate, runSearch, shouldRequireApproval, unitsForFile, upsertRegistry };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map