Compare commits

..
Author SHA1 Message Date
d9834a7a15 feat(ai): add reranker touchpoint to LiteLLM proxy recipe (takeover of #2455)
LiteLLM normalizes Cohere/Voyage/Jina rerank backends to the wire shape
gateway.rerank() already speaks, so a reranker touchpoint on the litellm
recipe makes any proxied rerank model reachable via
`search.reranker.model litellm:<model>` with no adapter.

Repairs from the original PR:
- path is the LEAF '/rerank' (not '/v1/rerank'): LiteLLM serves both
  /rerank and /v1/rerank, and the recipe's setup_hint allows
  LITELLM_BASE_URL with or without the /v1 suffix — pinning '/v1/rerank'
  doubled to /v1/v1/rerank (404) on /v1-suffixed bases.
- setup_hint appends the rerank guidance to master's current line instead
  of replacing it with a stale pre-/v1-suffix version.
- cost_per_1m_tokens_usd stays undefined (pricing-unknown), matching the
  recipe's embedding/chat touchpoints and budget-tracker's deliberate
  litellm exclusion from the free-provider sets (a proxy can front a paid
  provider; the touchpoint field isn't consumed by rerank pricing anyway).

Test drives gateway.rerank()'s real URL builder via the stubbed transport
for both base-URL forms; the /v1-suffixed case fails with the original
PR's path.

Co-authored-by: ozp <ozp@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-21 14:28:27 -07:00
13 changed files with 133 additions and 165 deletions
+31 -1
View File
@@ -56,6 +56,36 @@ export const litellmProxy: Recipe = {
cost_per_1m_output_usd: undefined,
price_last_verified: '2026-06-14',
},
// LiteLLM normalizes Cohere / Voyage / Jina / etc. rerank backends to the
// same wire shape gbrain's gateway.rerank() already speaks (the
// ZeroEntropy/llama.cpp contract):
// { model, query, documents, top_n } → { results: [{ index, relevance_score }] }
// So any rerank model the user registers in their LiteLLM config is
// reachable via `gbrain config set search.reranker.model litellm:<model>`
// with no request/response adapter — same as embeddings ride the proxy.
reranker: {
models: [], // user-provided; whatever rerank models the proxy serves
// No canonical default — the proxy defines its own model ids. The user
// sets search.reranker.model explicitly (mirrors the embedding
// touchpoint's user_provided_models contract).
default_model: '',
// The proxied backend bills (Cohere/Voyage/…); pricing-unknown is the
// honest state — same stance as this recipe's embedding/chat
// touchpoints and budget-tracker's deliberate litellm exclusion from
// the free-provider sets.
cost_per_1m_tokens_usd: undefined,
price_last_verified: '2026-06-27',
max_payload_bytes: 5_000_000,
// LEAF path only (matches llama-server-reranker's convention). LiteLLM
// serves both `/rerank` and `/v1/rerank`, and LITELLM_BASE_URL may be
// set with or without the `/v1` suffix (the setup_hint allows both), so
// the leaf form yields a valid route either way:
// http://localhost:4000 + /rerank → /rerank ✓
// http://localhost:4000/v1 + /rerank → /v1/rerank ✓
// Pinning '/v1/rerank' here would double to /v1/v1/rerank → 404 on
// /v1-suffixed bases.
path: '/rerank',
},
},
setup_hint: 'Run LiteLLM (https://docs.litellm.ai) in front of any provider; set LITELLM_BASE_URL (include the /v1 suffix if your proxy serves the OpenAI route there, e.g. http://localhost:4000/v1) + pass --embedding-model litellm:<model> and --embedding-dimensions <N>.',
setup_hint: 'Run LiteLLM (https://docs.litellm.ai) in front of any provider; set LITELLM_BASE_URL (include the /v1 suffix if your proxy serves the OpenAI route there, e.g. http://localhost:4000/v1) + pass --embedding-model litellm:<model> and --embedding-dimensions <N>. For rerank: register a rerank model in LiteLLM and set search.reranker.model litellm:<model-name>.',
};
-4
View File
@@ -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',
-2
View File
@@ -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;
/**
+5 -8
View File
@@ -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, {
-48
View File
@@ -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;
}
+7 -14
View File
@@ -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>> = [];
+1 -5
View File
@@ -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
+1 -5
View File
@@ -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 +
-8
View File
@@ -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.
-11
View File
@@ -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
+88
View File
@@ -0,0 +1,88 @@
/**
* litellm-proxy reranker touchpoint smoke.
*
* Sibling of recipe-llama-server-reranker.test.ts. Pins the reranker
* touchpoint on the LiteLLM proxy recipe so:
* - the touchpoint exists with the LEAF '/rerank' path (LiteLLM serves both
* /rerank and /v1/rerank, so the leaf form is valid whether or not the
* user's LITELLM_BASE_URL carries the /v1 suffix the setup_hint allows)
* - a /v1-suffixed base URL does NOT produce /v1/v1/rerank (the original
* community PR pinned '/v1/rerank' which 404s on /v1-suffixed bases)
* - models: [] (user-provided; proxy defines the model ids)
* - pricing stays undefined (proxy can front a paid provider — same honest
* pricing-unknown stance as the embedding/chat touchpoints)
*
* The gateway.rerank() URL tests drive the real URL builder via the stubbed
* transport (same seam as test/ai/rerank.test.ts).
*/
import { describe, expect, test, afterEach } from 'bun:test';
import { getRecipe } from '../../src/core/ai/recipes/index.ts';
import {
configureGateway,
resetGateway,
rerank,
__setRerankTransportForTests,
} from '../../src/core/ai/gateway.ts';
afterEach(() => {
__setRerankTransportForTests(null);
resetGateway();
});
describe('recipe: litellm reranker touchpoint', () => {
test('declares reranker touchpoint with leaf /rerank path', () => {
const r = getRecipe('litellm')!;
const tp = r.touchpoints.reranker;
expect(tp).toBeDefined();
expect(tp!.path).toBe('/rerank');
expect(tp!.max_payload_bytes).toBe(5_000_000);
});
test('reranker touchpoint uses empty models[] for user-provided model ids', () => {
const r = getRecipe('litellm')!;
expect(r.touchpoints.reranker!.models).toEqual([]);
});
test('pricing stays undefined — proxy can front a paid provider', () => {
const r = getRecipe('litellm')!;
expect(r.touchpoints.reranker!.cost_per_1m_tokens_usd).toBeUndefined();
});
test('setup_hint keeps the /v1-suffix guidance AND mentions rerank', () => {
const r = getRecipe('litellm')!;
expect(r.setup_hint).toMatch(/\/v1 suffix/);
expect(r.setup_hint).toMatch(/search\.reranker\.model litellm:/);
});
});
describe('gateway.rerank() URL via litellm recipe', () => {
async function capturedRerankUrl(baseUrl?: string): Promise<string> {
configureGateway({
reranker_model: 'litellm:my-reranker',
env: {},
...(baseUrl ? { base_urls: { litellm: baseUrl } } : {}),
});
let capturedUrl = '';
__setRerankTransportForTests(async (url) => {
capturedUrl = url;
return new Response(
JSON.stringify({ results: [{ index: 0, relevance_score: 0.9 }] }),
{ status: 200, headers: { 'content-type': 'application/json' } },
);
});
await rerank({ query: 'q', documents: ['d'] });
return capturedUrl;
}
test('default base (no /v1 suffix) → /rerank', async () => {
const url = await capturedRerankUrl();
expect(url).toBe('http://localhost:4000/rerank');
});
test('/v1-suffixed base → /v1/rerank, NOT /v1/v1/rerank', async () => {
const url = await capturedRerankUrl('http://localhost:4000/v1');
expect(url).toBe('http://localhost:4000/v1/rerank');
expect(url).not.toContain('/v1/v1/');
});
});
-16
View File
@@ -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', () => {
-43
View File
@@ -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();
});
});