mirror of
https://github.com/garrytan/gbrain.git
synced 2026-08-16 18:02:30 +00:00
Compare commits
2
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
89579780e0 | ||
|
|
cf2deedfc6 |
+14
-1
@@ -1,5 +1,5 @@
|
||||
import type { BrainEngine } from '../core/engine.ts';
|
||||
import { embedBatch, currentEmbeddingSignature } from '../core/embedding.ts';
|
||||
import { embedBatch, currentEmbeddingSignature, resolveEmbeddingModelLabel } from '../core/embedding.ts';
|
||||
import type { ChunkInput } from '../core/types.ts';
|
||||
import { chunkText } from '../core/chunkers/recursive.ts';
|
||||
import { createProgress, type ProgressReporter } from '../core/progress.ts';
|
||||
@@ -581,11 +581,16 @@ async function embedPage(
|
||||
for (let j = 0; j < toEmbed.length; j++) {
|
||||
embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
|
||||
}
|
||||
// #1717: label each (re)embedded chunk with the model that actually
|
||||
// produced its vector. Preserved chunks (not re-embedded this pass) keep
|
||||
// their existing model so a mixed-model page isn't relabeled wholesale.
|
||||
const embedModelLabel = resolveEmbeddingModelLabel();
|
||||
const updated: ChunkInput[] = chunks.map(c => ({
|
||||
chunk_index: c.chunk_index,
|
||||
chunk_text: c.chunk_text,
|
||||
chunk_source: c.chunk_source,
|
||||
embedding: embeddingMap.get(c.chunk_index),
|
||||
model: embeddingMap.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
|
||||
token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
|
||||
}));
|
||||
|
||||
@@ -717,12 +722,16 @@ async function embedAll(
|
||||
for (let j = 0; j < toEmbed.length; j++) {
|
||||
embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
|
||||
}
|
||||
// #1717: stamp the resolved embedding model on (re)embedded chunks;
|
||||
// preserve the existing model on chunks left untouched.
|
||||
const embedModelLabel = resolveEmbeddingModelLabel();
|
||||
// Preserve ALL chunks, only update embeddings for stale ones
|
||||
const updated: ChunkInput[] = chunks.map(c => ({
|
||||
chunk_index: c.chunk_index,
|
||||
chunk_text: c.chunk_text,
|
||||
chunk_source: c.chunk_source,
|
||||
embedding: embeddingMap.get(c.chunk_index) ?? undefined,
|
||||
model: embeddingMap.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
|
||||
token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
|
||||
}));
|
||||
await observed(pacer, () => engine.upsertChunks(page.slug, updated, pageOpts));
|
||||
@@ -1012,11 +1021,15 @@ async function embedAllStale(
|
||||
for (let j = 0; j < stale.length; j++) {
|
||||
staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
|
||||
}
|
||||
// #1717: label the re-embedded (stale) chunks with the resolved
|
||||
// model; preserve the existing model on the non-stale chunks.
|
||||
const embedModelLabel = resolveEmbeddingModelLabel();
|
||||
const merged: ChunkInput[] = existing.map(c => ({
|
||||
chunk_index: c.chunk_index,
|
||||
chunk_text: c.chunk_text,
|
||||
chunk_source: c.chunk_source,
|
||||
embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
|
||||
model: staleIdxToEmbedding.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
|
||||
token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
|
||||
}));
|
||||
await observed(pacer, () => engine.upsertChunks(slug, merged, { sourceId: keySourceId }));
|
||||
|
||||
@@ -908,10 +908,6 @@ export const KNOWN_CONFIG_KEYS: readonly string[] = [
|
||||
'facts.extraction_model',
|
||||
// #2113: output-token cap for the per-turn facts extractor (default 4000).
|
||||
'facts.extraction_max_tokens',
|
||||
// Owner opt-in: let the local stdio MCP pipe read this owner's private facts
|
||||
// (find_trajectory / recall). Default off; HTTP transport ignores it. See
|
||||
// src/core/facts/reader-trust.ts.
|
||||
'facts.trust_local_reads',
|
||||
// Dream cycle config
|
||||
'dream.synthesize.session_corpus_dir',
|
||||
'dream.synthesize.meeting_transcripts_dir',
|
||||
|
||||
@@ -20,6 +20,7 @@
|
||||
import type { BrainEngine } from './engine.ts';
|
||||
import type { ChunkInput } from './types.ts';
|
||||
import { embedBatchWithBackoff } from '../commands/embed.ts';
|
||||
import { resolveEmbeddingModelLabel } from './embedding.ts';
|
||||
import { type DbPacer, createNoopPacer, observed } from './db-pacer.ts';
|
||||
import { AbortError } from './abort-check.ts';
|
||||
|
||||
@@ -200,11 +201,17 @@ export async function embedStaleForSource(
|
||||
for (let j = 0; j < stale.length; j++) {
|
||||
staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
|
||||
}
|
||||
// #1717: label re-embedded chunks with the model that produced the
|
||||
// vector; preserved chunks keep their existing model. Without this,
|
||||
// upsertChunks falls back to DEFAULT_EMBEDDING_MODEL for every chunk
|
||||
// (the same mislabel the embed.ts paths fixed).
|
||||
const embedModelLabel = resolveEmbeddingModelLabel();
|
||||
const merged: ChunkInput[] = existing.map((c) => ({
|
||||
chunk_index: c.chunk_index,
|
||||
chunk_text: c.chunk_text,
|
||||
chunk_source: c.chunk_source,
|
||||
embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
|
||||
model: staleIdxToEmbedding.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
|
||||
token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
|
||||
// Carry through per-chunk metadata. upsertChunks writes these as
|
||||
// EXCLUDED.<col> (not COALESCE), so omitting them here resets image
|
||||
|
||||
@@ -113,6 +113,21 @@ export async function embedBatch(
|
||||
return results;
|
||||
}
|
||||
|
||||
/**
|
||||
* Resolve the embedding model label (`provider:model`) to stamp onto
|
||||
* `content_chunks.model`, so each chunk records the model that actually
|
||||
* produced its vector instead of the engine's hardcoded default (#1717).
|
||||
* Returns undefined if the gateway is unconfigured; callers then fall back
|
||||
* to the chunk's existing model rather than mislabeling it.
|
||||
*/
|
||||
export function resolveEmbeddingModelLabel(): string | undefined {
|
||||
try {
|
||||
return gatewayGetModel();
|
||||
} catch {
|
||||
return undefined;
|
||||
}
|
||||
}
|
||||
|
||||
/** Currently-configured embedding model (short form without provider prefix). */
|
||||
export function getEmbeddingModelName(): string {
|
||||
return gatewayGetModel().split(':').slice(1).join(':') || 'text-embedding-3-large';
|
||||
|
||||
@@ -591,8 +591,6 @@ export interface TrajectoryOpts {
|
||||
sourceIds?: string[];
|
||||
/** When true, filters to visibility='world' only. Set by MCP layer from ctx.remote. */
|
||||
remote?: boolean;
|
||||
/** Owner opt-in: read private facts despite `remote`. See facts/reader-trust.ts. */
|
||||
trustedFactReads?: boolean;
|
||||
/** Metric filter. When set, only facts with this canonical metric label participate. */
|
||||
metric?: string;
|
||||
/**
|
||||
|
||||
@@ -17,7 +17,6 @@
|
||||
import type { OperationContext } from './../operations.ts';
|
||||
import type { FactRow } from './../engine.ts';
|
||||
import { effectiveConfidence } from './decay.ts';
|
||||
import { readableFactVisibilities } from './reader-trust.ts';
|
||||
|
||||
const DEFAULT_TTL_MS = 30_000;
|
||||
const DEFAULT_TOP_K = 10;
|
||||
@@ -51,13 +50,7 @@ export async function getBrainHotMemoryMeta(
|
||||
const sessionId = (ctx as { source_session?: string }).source_session
|
||||
?? null;
|
||||
const allowListHash = hashAllowList(ctx.takesHoldersAllowList);
|
||||
// Visibility tier: untrusted remote → world-only; trusted local +
|
||||
// owner-trusted reads → all rows. Folded into the cache key (the header's
|
||||
// "cache entries don't bleed across tiers" invariant): trustedFactReads is
|
||||
// re-read from config per call, so a mid-session opt-out must not keep
|
||||
// serving a private-inclusive cached payload for the TTL window.
|
||||
const visibility = readableFactVisibilities(ctx);
|
||||
const cacheKey = `${sourceId}::${sessionId ?? '_'}::${allowListHash}::${visibility ? 'world' : 'all'}`;
|
||||
const cacheKey = `${sourceId}::${sessionId ?? '_'}::${allowListHash}`;
|
||||
|
||||
const ttl = Math.max(1000, opts.ttlMs ?? DEFAULT_TTL_MS);
|
||||
const topK = Math.max(1, Math.min(opts.topK ?? DEFAULT_TOP_K, 25));
|
||||
@@ -68,6 +61,10 @@ export async function getBrainHotMemoryMeta(
|
||||
return cached.payload;
|
||||
}
|
||||
|
||||
// Build a fresh payload. Visibility tier: remote → world-only;
|
||||
// local → all rows.
|
||||
const visibility = ctx.remote === false ? undefined : ['world'] as ('world' | 'private')[];
|
||||
|
||||
let rows: FactRow[] = [];
|
||||
if (sessionId) {
|
||||
rows = await ctx.engine.listFactsBySession(sourceId, sessionId, {
|
||||
|
||||
@@ -1,48 +0,0 @@
|
||||
/**
|
||||
* Fact-read visibility trust.
|
||||
*
|
||||
* Fact rows are tagged `private` | `world`. Remote/untrusted callers
|
||||
* (`remote === true`) see only `world` rows — the posture that keeps a
|
||||
* published or HTTP-served brain from leaking private claims to strangers.
|
||||
*
|
||||
* But the stdio MCP server is an unauthenticated LOCAL pipe: on a single-owner
|
||||
* machine the caller IS the owner, yet it still defaults `remote: true` for
|
||||
* safety, so the owner's own agent is denied the owner's own private facts
|
||||
* (e.g. `find_trajectory` returns empty over MCP even though the facts exist).
|
||||
*
|
||||
* `trustedFactReads` is a narrow, READ-ONLY trust elevation, deliberately
|
||||
* DECOUPLED from `remote` so every other remote protection — file_upload
|
||||
* confinement, source isolation, fence stripping, takes-holder scoping — stays
|
||||
* fully in force. The stdio MCP server sets it ONLY when the brain owner opts
|
||||
* in via the `facts.trust_local_reads` config (default off). The HTTP/published
|
||||
* transport never sets it, so a served brain stays world-only regardless.
|
||||
*/
|
||||
export interface FactReaderTrust {
|
||||
/**
|
||||
* Mirrors OperationContext.remote. FAIL-CLOSED: anything not strictly
|
||||
* `false` is treated as remote/untrusted (CLAUDE.md trust invariant).
|
||||
*/
|
||||
remote?: boolean;
|
||||
/** Owner opt-in: this remote caller may read private facts. */
|
||||
trustedFactReads?: boolean;
|
||||
}
|
||||
|
||||
/**
|
||||
* True when the reader is restricted to `visibility = 'world'` rows.
|
||||
* Fail-closed: an unset/undefined `remote` is untrusted — only an explicit
|
||||
* `remote: false` (trusted local CLI) or an explicit owner opt-in
|
||||
* (`trustedFactReads: true`) reads private rows.
|
||||
*/
|
||||
export function factsWorldOnly(t: FactReaderTrust): boolean {
|
||||
return t.remote !== false && t.trustedFactReads !== true;
|
||||
}
|
||||
|
||||
/**
|
||||
* Visibility filter for list-style fact reads: `['world']` when the reader is
|
||||
* world-only, `undefined` (no filter — all rows) when it is trusted.
|
||||
*/
|
||||
export function readableFactVisibilities(
|
||||
t: FactReaderTrust,
|
||||
): ('private' | 'world')[] | undefined {
|
||||
return factsWorldOnly(t) ? ['world'] : undefined;
|
||||
}
|
||||
+11
-1
@@ -8,7 +8,7 @@ import { chunkText } from './chunkers/recursive.ts';
|
||||
import { chunkCodeText, chunkCodeTextFull, detectCodeLanguage, CHUNKER_VERSION } from './chunkers/code.ts';
|
||||
import { findChunkForOffset } from './chunkers/edge-extractor.ts';
|
||||
import { extractCodeRefs, imageOfCandidates } from './link-extraction.ts';
|
||||
import { embedBatch, embedMultimodal, currentEmbeddingSignature } from './embedding.ts';
|
||||
import { embedBatch, embedMultimodal, currentEmbeddingSignature, resolveEmbeddingModelLabel } from './embedding.ts';
|
||||
import { slugifyPath, slugifyCodePath, isCodeFilePath } from './sync.ts';
|
||||
import type { ChunkInput, PageInput, PageType } from './types.ts';
|
||||
import { computeEffectiveDate } from './effective-date.ts';
|
||||
@@ -716,8 +716,12 @@ export async function importFromContent(
|
||||
? chunks.map((c) => wrapChunkForEmbedding(c.chunk_text, prefix, c.chunk_source))
|
||||
: chunks.map((c) => c.chunk_text);
|
||||
const embeddings = await embedBatch(wrappedTexts);
|
||||
// #1717: label each chunk with the model that actually produced its
|
||||
// vector, not the engine's hardcoded default.
|
||||
const embedModelLabel = resolveEmbeddingModelLabel();
|
||||
for (let i = 0; i < chunks.length; i++) {
|
||||
chunks[i].embedding = embeddings[i];
|
||||
if (embedModelLabel) chunks[i].model = embedModelLabel;
|
||||
// token_count tracks the wrapped string length so cost reporting
|
||||
// reflects what we actually sent to the embedder.
|
||||
chunks[i].token_count = Math.ceil(wrappedTexts[i].length / 4);
|
||||
@@ -1141,7 +1145,10 @@ export async function importCodeFile(
|
||||
const matched = existingByKey.get(key);
|
||||
if (matched && matched.embedding) {
|
||||
// Reuse the existing embedding verbatim. No API call, no cost.
|
||||
// #1717: carry the existing model label along with the reused vector
|
||||
// so the upsert doesn't relabel it with the engine default.
|
||||
chunks[i]!.embedding = matched.embedding as Float32Array;
|
||||
chunks[i]!.model = matched.model ?? undefined;
|
||||
chunks[i]!.token_count = matched.token_count ?? undefined;
|
||||
} else {
|
||||
needsEmbedIndexes.push(i);
|
||||
@@ -1153,9 +1160,12 @@ export async function importCodeFile(
|
||||
try {
|
||||
const textsToEmbed = needsEmbedIndexes.map((i) => chunks[i]!.chunk_text);
|
||||
const embeddings = await embedBatch(textsToEmbed);
|
||||
// #1717: stamp the model that produced these vectors.
|
||||
const embedModelLabel = resolveEmbeddingModelLabel();
|
||||
for (let j = 0; j < needsEmbedIndexes.length; j++) {
|
||||
const i = needsEmbedIndexes[j]!;
|
||||
chunks[i]!.embedding = embeddings[j]!;
|
||||
if (embedModelLabel) chunks[i]!.model = embedModelLabel;
|
||||
chunks[i]!.token_count = Math.ceil(chunks[i]!.chunk_text.length / 4);
|
||||
}
|
||||
} catch (e: unknown) {
|
||||
|
||||
+7
-14
@@ -18,7 +18,6 @@ import { captureEvalCandidate, isEvalCaptureEnabled, isEvalScrubEnabled } from '
|
||||
import type { HybridSearchMeta } from './types.ts';
|
||||
import { extractPageLinks, isAutoLinkEnabled, isAutoTimelineEnabled, isGlobalBasenameEnabled, parseTimelineEntries, makeResolver, type UnresolvedFrontmatterRef } from './link-extraction.ts';
|
||||
import { isFactsBackstopEligible } from './facts/eligibility.ts';
|
||||
import { readableFactVisibilities } from './facts/reader-trust.ts';
|
||||
import { stripTakesFence } from './takes-fence.ts';
|
||||
import { stripFactsFence } from './facts-fence.ts';
|
||||
import { getContentFlag } from './quarantine.ts';
|
||||
@@ -333,15 +332,6 @@ export interface OperationContext {
|
||||
* remote/untrusted (defense in depth in case the type is bypassed via cast).
|
||||
*/
|
||||
remote: boolean;
|
||||
/**
|
||||
* Owner opt-in (`facts.trust_local_reads`): allow this remote caller to read
|
||||
* `private` facts. A NARROW, read-only trust elevation decoupled from
|
||||
* `remote` — every other remote protection (file confinement, source
|
||||
* isolation, fence stripping, takes scoping) stays in force. Set ONLY by the
|
||||
* stdio MCP server when the config is on; the HTTP transport never sets it.
|
||||
* Consulted via `src/core/facts/reader-trust.ts`.
|
||||
*/
|
||||
trustedFactReads?: boolean;
|
||||
/**
|
||||
* Subagent runtime context (v0.16+). Set by the subagent tool dispatcher when
|
||||
* dispatching an op as a tool call from an LLM loop. Used to enforce per-op
|
||||
@@ -3639,7 +3629,6 @@ const find_trajectory: Operation = {
|
||||
entitySlug: p.entity_slug,
|
||||
...scope,
|
||||
remote: ctx.remote === true,
|
||||
trustedFactReads: ctx.trustedFactReads === true,
|
||||
metric,
|
||||
kind,
|
||||
since,
|
||||
@@ -4000,9 +3989,13 @@ const recall: Operation = {
|
||||
const includeExpired = p.include_expired === true;
|
||||
const grep = typeof p.grep === 'string' ? p.grep.toLowerCase() : null;
|
||||
|
||||
// Visibility filter: world-only for untrusted remote callers; all rows for
|
||||
// trusted local CLI and owner-trusted reads (facts.trust_local_reads).
|
||||
const visibility = readableFactVisibilities(ctx);
|
||||
// Visibility filter: remote callers see world-only unless their token
|
||||
// grants elevated visibility (future-proofing; v0.31 ships world-only
|
||||
// for remote, all for local CLI).
|
||||
const visibility =
|
||||
ctx.remote === false
|
||||
? undefined
|
||||
: ['world'] as ('private' | 'world')[];
|
||||
|
||||
let rows: Awaited<ReturnType<typeof ctx.engine.listFactsByEntity>> = [];
|
||||
|
||||
|
||||
@@ -18,7 +18,6 @@ import type {
|
||||
} from './engine.ts';
|
||||
import { MAX_SEARCH_LIMIT, clampSearchLimit } from './engine.ts';
|
||||
import { withRetry, BULK_RETRY_OPTS, resolveBulkRetryOpts, computeNextDelay, type BatchAuditSite } from './retry.ts';
|
||||
import { factsWorldOnly } from './facts/reader-trust.ts';
|
||||
import { logBatchRetry as auditLogBatchRetry, logBatchExhausted as auditLogBatchExhausted } from './audit/batch-retry-audit.ts';
|
||||
import { runMigrations } from './migrate.ts';
|
||||
import { PGLITE_SCHEMA_SQL, getPGLiteSchema } from './pglite-schema.ts';
|
||||
@@ -4319,10 +4318,7 @@ export class PGLiteEngine implements BrainEngine {
|
||||
const useArray = Array.isArray(opts.sourceIds) && opts.sourceIds.length > 0;
|
||||
const sourceIds = useArray ? opts.sourceIds! : null;
|
||||
const sourceId = opts.sourceId ?? 'default';
|
||||
// Direct-engine contract: unset `remote` here means a trusted in-process
|
||||
// caller (CLI/tests) — the fail-closed default lives in the op layer, which
|
||||
// always passes explicit booleans. Normalize before the fail-closed helper.
|
||||
const remoteFilter = factsWorldOnly({ remote: opts.remote === true, trustedFactReads: opts.trustedFactReads === true });
|
||||
const remoteFilter = opts.remote === true;
|
||||
|
||||
// Build SQL dynamically. PGLite uses $N positional params; we
|
||||
// assemble the WHERE clauses + params array in tandem to keep them
|
||||
|
||||
@@ -14,7 +14,6 @@ import type {
|
||||
SourceRow,
|
||||
} from './engine.ts';
|
||||
import { withRetry, BULK_RETRY_OPTS, resolveBulkRetryOpts, computeNextDelay, type BatchAuditSite } from './retry.ts';
|
||||
import { factsWorldOnly } from './facts/reader-trust.ts';
|
||||
import { logBatchRetry as auditLogBatchRetry, logBatchExhausted as auditLogBatchExhausted } from './audit/batch-retry-audit.ts';
|
||||
import type {
|
||||
DomainBankSampleOpts, CorpusSampleOpts, DomainBankRow,
|
||||
@@ -4533,10 +4532,7 @@ export class PostgresEngine implements BrainEngine {
|
||||
const useArray = Array.isArray(opts.sourceIds) && opts.sourceIds.length > 0;
|
||||
const sourceIds = useArray ? opts.sourceIds! : null;
|
||||
const sourceId = opts.sourceId ?? 'default';
|
||||
// Direct-engine contract: unset `remote` here means a trusted in-process
|
||||
// caller (CLI/tests) — the fail-closed default lives in the op layer, which
|
||||
// always passes explicit booleans. Normalize before the fail-closed helper.
|
||||
const remoteFilter = factsWorldOnly({ remote: opts.remote === true, trustedFactReads: opts.trustedFactReads === true });
|
||||
const remoteFilter = opts.remote === true;
|
||||
|
||||
// Source-scope predicate: array path (federated) wins over scalar.
|
||||
// Engine.ts contract: returns chronological points; regressions +
|
||||
|
||||
@@ -30,13 +30,6 @@ export interface ToolResult {
|
||||
export interface DispatchOpts {
|
||||
/** Defaults to true (remote/untrusted). Local CLI callers (`gbrain call`) pass false. */
|
||||
remote?: boolean;
|
||||
/**
|
||||
* Owner opt-in (`facts.trust_local_reads`): let this remote caller read
|
||||
* `private` facts. Set ONLY by the stdio MCP server; the HTTP transport
|
||||
* leaves it unset so a served brain stays world-only. See
|
||||
* `src/core/facts/reader-trust.ts`.
|
||||
*/
|
||||
trustedFactReads?: boolean;
|
||||
/** Override the default stderr logger (e.g. CLI uses console.* directly). */
|
||||
logger?: OperationContext['logger'];
|
||||
/**
|
||||
@@ -210,7 +203,6 @@ export function buildOperationContext(
|
||||
logger: opts.logger || stderrLogger,
|
||||
dryRun: !!params.dry_run,
|
||||
remote: opts.remote ?? true,
|
||||
trustedFactReads: opts.trustedFactReads === true,
|
||||
takesHoldersAllowList: opts.takesHoldersAllowList,
|
||||
// v0.34 D4: sourceId is REQUIRED at the type level. Auto-fill 'default'
|
||||
// for single-source brains and any caller who didn't resolve a sourceId.
|
||||
|
||||
@@ -35,16 +35,6 @@ export async function startMcpServer(engine: BrainEngine) {
|
||||
// shape and cast through `any` (the SDK accepts it via the ServerResult union).
|
||||
server.setRequestHandler(CallToolRequestSchema, async (request: any): Promise<any> => {
|
||||
const { name, arguments: params } = request.params;
|
||||
// Owner opt-in: the stdio pipe is local + unauthenticated, so on a
|
||||
// single-owner machine its caller is the owner. When facts.trust_local_reads
|
||||
// is on, let fact reads (find_trajectory / recall) see this owner's own
|
||||
// private facts. Narrow + read-only — every other remote protection stays
|
||||
// on (remote stays true). HTTP transport never sets this. Best-effort: a
|
||||
// config read blip falls back to the safe world-only default.
|
||||
let trustedFactReads = false;
|
||||
try {
|
||||
trustedFactReads = (await engine.getConfig('facts.trust_local_reads')) === 'true';
|
||||
} catch { /* keep world-only default */ }
|
||||
// v0.28: stdio MCP has no per-token auth (local pipe). Default the
|
||||
// takes-holder allow-list to ['world'] so agent-facing callers don't
|
||||
// see private hunches via takes_list / takes_search / query. Operators
|
||||
@@ -52,7 +42,6 @@ export async function startMcpServer(engine: BrainEngine) {
|
||||
// `gbrain call <op>` (sets remote=false in src/cli.ts).
|
||||
return dispatchToolCall(engine, name, params, {
|
||||
remote: true,
|
||||
trustedFactReads,
|
||||
takesHoldersAllowList: ['world'],
|
||||
// v0.31: source defaults to 'default' for stdio (no per-token scope).
|
||||
// Operators who want a different source on stdio MCP should set
|
||||
|
||||
@@ -15,6 +15,7 @@ import { describe, test, expect, beforeAll, afterAll, beforeEach } from 'bun:tes
|
||||
import { PGLiteEngine } from '../src/core/pglite-engine.ts';
|
||||
import { resetPgliteState } from './helpers/reset-pglite.ts';
|
||||
import { embedStaleForSource } from '../src/core/embed-stale.ts';
|
||||
import { configureGateway, resetGateway } from '../src/core/ai/gateway.ts';
|
||||
import type { ChunkInput } from '../src/core/types.ts';
|
||||
|
||||
let engine: PGLiteEngine;
|
||||
@@ -276,4 +277,49 @@ describe('embedStaleForSource', () => {
|
||||
// The stale text row actually got its embedding.
|
||||
expect(txtRow.embedded_at).not.toBeNull();
|
||||
});
|
||||
|
||||
// #1717: the backfill path must label re-embedded chunks with the model
|
||||
// that produced the vector, and preserve the existing label on chunks it
|
||||
// did not touch (before the fix, both were reset to the engine default).
|
||||
test('labels re-embedded chunks with the gateway model, preserves untouched labels (#1717)', async () => {
|
||||
configureGateway({
|
||||
embedding_model: 'openai:text-embedding-3-large',
|
||||
env: { OPENAI_API_KEY: 'sk-test-embed-stale-1717' },
|
||||
});
|
||||
try {
|
||||
await engine.putPage('notes/model-label', {
|
||||
type: 'note',
|
||||
title: 'model-label',
|
||||
compiled_truth: '# model-label\n\nseeded',
|
||||
});
|
||||
await engine.upsertChunks('notes/model-label', [
|
||||
{
|
||||
chunk_index: 0,
|
||||
chunk_text: 'already embedded elsewhere',
|
||||
chunk_source: 'compiled_truth',
|
||||
embedding: new Float32Array(1536).fill(0.01),
|
||||
model: 'voyage:voyage-3',
|
||||
token_count: 4,
|
||||
},
|
||||
{
|
||||
chunk_index: 1,
|
||||
chunk_text: 'stale chunk needing embed',
|
||||
chunk_source: 'compiled_truth',
|
||||
token_count: 5,
|
||||
embedding: undefined, // stale
|
||||
},
|
||||
]);
|
||||
|
||||
const result = await embedStaleForSource(engine, 'default', { embedFn: fakeEmbedFn });
|
||||
expect(result.embedded).toBe(1);
|
||||
|
||||
const after = await engine.getChunks('notes/model-label');
|
||||
const preserved = after.find((c) => c.chunk_index === 0)!;
|
||||
const reembedded = after.find((c) => c.chunk_index === 1)!;
|
||||
expect(reembedded.model).toBe('openai:text-embedding-3-large');
|
||||
expect(preserved.model).toBe('voyage:voyage-3');
|
||||
} finally {
|
||||
resetGateway();
|
||||
}
|
||||
});
|
||||
});
|
||||
|
||||
@@ -37,6 +37,8 @@ mock.module('../src/core/embedding.ts', () => ({
|
||||
// setPageEmbeddingSignature / invalidateStaleSignatureEmbeddings resolve to
|
||||
// null via the Proxy default, so the signature value is inert here.
|
||||
currentEmbeddingSignature: () => 'test:model:1536',
|
||||
// #1717: embed paths stamp this label on (re)embedded chunks.
|
||||
resolveEmbeddingModelLabel: () => 'openai:text-embedding-3-large',
|
||||
}));
|
||||
|
||||
// Import AFTER mocking.
|
||||
@@ -803,3 +805,34 @@ describe('embedAllStale --source threading (D7)', () => {
|
||||
expect((firstCallOpts as { sourceId?: string }).sourceId).toBe('media-corpus');
|
||||
});
|
||||
});
|
||||
|
||||
// #1717: content_chunks.model must record the model that actually produced
|
||||
// each vector, not the gateway/engine default.
|
||||
describe('content_chunks.model labeling (#1717)', () => {
|
||||
test('stamps the resolved embedding model on re-embedded chunks, preserves it on untouched chunks', async () => {
|
||||
let upserted: any[] | undefined;
|
||||
// Chunk 0 is stale (no embedded_at) → gets re-embedded this pass.
|
||||
// Chunk 1 is already embedded with a DIFFERENT model → must be preserved,
|
||||
// not relabeled to the current model.
|
||||
const chunks = [
|
||||
{ chunk_index: 0, chunk_text: 'a', chunk_source: 'compiled_truth', embedded_at: null, model: 'zeroentropyai:zembed-1', token_count: 1 },
|
||||
{ chunk_index: 1, chunk_text: 'b', chunk_source: 'compiled_truth', embedded_at: '2026-01-01', embedding: new Float32Array(1536), model: 'voyage:voyage-3', token_count: 1 },
|
||||
];
|
||||
const engine = mockEngine({
|
||||
getPage: async () => ({ slug: 'notes/x', compiled_truth: 'a', timeline: '', source_id: 'default' }),
|
||||
getChunks: async () => chunks,
|
||||
upsertChunks: async (_slug: string, c: any[]) => { upserted = c; },
|
||||
setPageEmbeddingSignature: async () => null,
|
||||
});
|
||||
|
||||
await runEmbedCore(engine, { slugs: ['notes/x'] });
|
||||
|
||||
expect(upserted).toBeDefined();
|
||||
const byIdx = Object.fromEntries(upserted!.map(c => [c.chunk_index, c]));
|
||||
// Re-embedded chunk carries the model that produced its vector (was
|
||||
// mislabeled with the default before the fix).
|
||||
expect(byIdx[0].model).toBe('openai:text-embedding-3-large');
|
||||
// Untouched chunk keeps its original model — no wholesale relabel.
|
||||
expect(byIdx[1].model).toBe('voyage:voyage-3');
|
||||
});
|
||||
});
|
||||
|
||||
@@ -165,22 +165,6 @@ describe('findTrajectory — visibility filter (D-CDX-1 / R6)', () => {
|
||||
const all = await engine.findTrajectory({ entitySlug: 'traj-vis-default' });
|
||||
expect(all.length).toBe(2);
|
||||
});
|
||||
|
||||
test('remote=true + trustedFactReads bypasses world-only (owner-trusted reads)', async () => {
|
||||
await insertTyped({ entity_slug: 'traj-vis-trusted', metric: 'mrr', value: 50000, visibility: 'private', valid_from: new Date('2026-01-15') });
|
||||
await insertTyped({ entity_slug: 'traj-vis-trusted', metric: 'mrr', value: 99999, visibility: 'world', valid_from: new Date('2026-04-12') });
|
||||
|
||||
// Untrusted remote: world only.
|
||||
const untrusted = await engine.findTrajectory({ entitySlug: 'traj-vis-trusted', remote: true });
|
||||
expect(untrusted.length).toBe(1);
|
||||
expect(untrusted[0].value).toBe(99999);
|
||||
|
||||
// Owner-trusted remote: sees the private point too. remote stays true — only
|
||||
// fact-read visibility is elevated.
|
||||
const trusted = await engine.findTrajectory({ entitySlug: 'traj-vis-trusted', remote: true, trustedFactReads: true });
|
||||
expect(trusted.length).toBe(2);
|
||||
expect(trusted.map(p => p.value).sort((a, b) => (a! - b!))).toEqual([50000, 99999]);
|
||||
});
|
||||
});
|
||||
|
||||
describe('findTrajectory — metric + since + until filters', () => {
|
||||
|
||||
@@ -1,43 +0,0 @@
|
||||
import { describe, test, expect } from 'bun:test';
|
||||
import {
|
||||
factsWorldOnly,
|
||||
readableFactVisibilities,
|
||||
} from '../src/core/facts/reader-trust.ts';
|
||||
|
||||
describe('factsWorldOnly', () => {
|
||||
test('FAIL-CLOSED: unset remote is untrusted (world-only)', () => {
|
||||
expect(factsWorldOnly({})).toBe(true);
|
||||
expect(factsWorldOnly({ remote: undefined })).toBe(true);
|
||||
expect(factsWorldOnly({ trustedFactReads: false })).toBe(true);
|
||||
});
|
||||
|
||||
test('explicit remote=false (trusted local CLI) sees all', () => {
|
||||
expect(factsWorldOnly({ remote: false })).toBe(false);
|
||||
expect(factsWorldOnly({ remote: false, trustedFactReads: false })).toBe(false);
|
||||
});
|
||||
|
||||
test('untrusted remote callers are world-only', () => {
|
||||
expect(factsWorldOnly({ remote: true })).toBe(true);
|
||||
expect(factsWorldOnly({ remote: true, trustedFactReads: false })).toBe(true);
|
||||
});
|
||||
|
||||
test('owner-trusted remote reads bypass the world-only filter', () => {
|
||||
expect(factsWorldOnly({ remote: true, trustedFactReads: true })).toBe(false);
|
||||
});
|
||||
|
||||
test('trustedFactReads is a no-op for an already-trusted local caller', () => {
|
||||
expect(factsWorldOnly({ remote: false, trustedFactReads: true })).toBe(false);
|
||||
});
|
||||
});
|
||||
|
||||
describe('readableFactVisibilities', () => {
|
||||
test("world-only readers get the ['world'] filter (incl. unset remote)", () => {
|
||||
expect(readableFactVisibilities({ remote: true })).toEqual(['world']);
|
||||
expect(readableFactVisibilities({})).toEqual(['world']);
|
||||
});
|
||||
|
||||
test('trusted readers get undefined (no filter — all rows)', () => {
|
||||
expect(readableFactVisibilities({ remote: false })).toBeUndefined();
|
||||
expect(readableFactVisibilities({ remote: true, trustedFactReads: true })).toBeUndefined();
|
||||
});
|
||||
});
|
||||
@@ -73,4 +73,21 @@ describe('importFromContent embedding_signature stamping (F1)', () => {
|
||||
await importFromContent(engine, 'concepts/unstamped', '# Unstamped\n\nbody content.', { noEmbed: true });
|
||||
expect(await signatureOf('concepts/unstamped')).toBeNull();
|
||||
});
|
||||
|
||||
// #1717: content_chunks.model must record the model that produced the
|
||||
// vector (the configured gateway model), not the engine's hardcoded
|
||||
// default. The gateway here is configured to openai:text-embedding-3-large,
|
||||
// which differs from DEFAULT_EMBEDDING_MODEL — so this fails without the
|
||||
// import-path model stamping.
|
||||
test('inline embed labels content_chunks.model with the configured model (#1717)', async () => {
|
||||
await importFromContent(engine, 'concepts/labeled', '# Labeled\n\nsome body content to chunk and embed.', {});
|
||||
const rows = await engine.executeRaw<{ model: string }>(
|
||||
`SELECT cc.model FROM content_chunks cc
|
||||
JOIN pages p ON p.id = cc.page_id
|
||||
WHERE p.slug = $1 AND p.source_id = 'default'`,
|
||||
['concepts/labeled'],
|
||||
);
|
||||
expect(rows.length).toBeGreaterThan(0);
|
||||
for (const r of rows) expect(r.model).toBe('openai:text-embedding-3-large');
|
||||
});
|
||||
});
|
||||
|
||||
Reference in New Issue
Block a user