fix(agent): make the recall reranker opt-in and use the word embedder in tests
The reference-set comparison showed the cross-encoder adds nothing to mem0's ranking and costs about 130 ms per search.
This commit is contained in:
@@ -78,10 +78,12 @@ export interface OrchestratorConfig {
|
||||
version?: string;
|
||||
skills?: string[];
|
||||
/**
|
||||
* W4.2: optional cross-encoder reranker injected for tests. When absent,
|
||||
* a lazy in-process reranker (Xenova/ms-marco-MiniLM-L-6-v2, ~22MB ONNX)
|
||||
* is created on first recall unless WAGGLE_RERANKER=0; creation failure
|
||||
* soft-fails to the store's own ordering.
|
||||
* W4.2: optional cross-encoder reranker injected for tests. When absent and
|
||||
* WAGGLE_RERANKER=1, a lazy in-process reranker (Xenova/ms-marco-MiniLM-L-6-v2,
|
||||
* ~22MB ONNX) is created on first recall; creation failure soft-fails to the
|
||||
* store's own ordering. Off by default: on the reference set it did not
|
||||
* improve mem0's ranking and cost about 130 ms per search
|
||||
* (docs/oss/research/memory-search-quality.md).
|
||||
*/
|
||||
reranker?: Reranker;
|
||||
/** Optional managed cache root for the lazy in-process reranker model. */
|
||||
@@ -366,14 +368,14 @@ export class Orchestrator {
|
||||
}
|
||||
|
||||
/**
|
||||
* W4.2/W4.5: lazy cross-encoder reranker — default on. Kill switch:
|
||||
* WAGGLE_RERANKER=0. First use downloads the ~22MB model (cached at the
|
||||
* configured managed path); creation failure (offline, OOM) memoizes
|
||||
* undefined: recall soft-fails to the store's ordering, never throws.
|
||||
* W4.2: lazy cross-encoder reranker, opt-in with WAGGLE_RERANKER=1. First use
|
||||
* downloads the ~22MB model (cached at the configured managed path); creation
|
||||
* failure (offline, OOM) memoizes undefined: recall soft-fails to the store's
|
||||
* ordering, never throws.
|
||||
*/
|
||||
private getReranker(): Promise<Reranker | undefined> {
|
||||
if (this.rerankerPromise) return this.rerankerPromise;
|
||||
if (process.env['WAGGLE_RERANKER'] === '0') {
|
||||
if (process.env['WAGGLE_RERANKER'] !== '1') {
|
||||
this.rerankerPromise = Promise.resolve(undefined);
|
||||
return this.rerankerPromise;
|
||||
}
|
||||
|
||||
@@ -28,14 +28,20 @@ if (!process.env.WAGGLE_PHASE5_CANARY_PCT) {
|
||||
process.env.WAGGLE_PHASE5_CANARY_PCT = '0';
|
||||
}
|
||||
|
||||
// W4.5: the cross-encoder reranker is default-ON in production (kill switch
|
||||
// WAGGLE_RERANKER=0). Tests pin it OFF — the real ONNX model is a ~22MB
|
||||
// download + per-call CPU inference, both unwanted in unit suites. Tests
|
||||
// that exercise reranking inject a mock Reranker via OrchestratorConfig.
|
||||
// The cross-encoder reranker is opt-in (WAGGLE_RERANKER=1). Tests pin it
|
||||
// OFF explicitly; tests that exercise reranking inject a mock Reranker via
|
||||
// OrchestratorConfig.
|
||||
if (!process.env.WAGGLE_RERANKER) {
|
||||
process.env.WAGGLE_RERANKER = '0';
|
||||
}
|
||||
|
||||
// Memory (ADR-0007): the sidecar embeds memories with an in-process model
|
||||
// that is a ~90MB download. Tests use the deterministic word embedder, which
|
||||
// needs no model and no network.
|
||||
if (!process.env.WAGGLE_MEMORY_EMBEDDER) {
|
||||
process.env.WAGGLE_MEMORY_EMBEDDER = 'words';
|
||||
}
|
||||
|
||||
// W4 (2026-06-11): the PromptAssembler is default-ON in production (kill
|
||||
// switch WAGGLE_PROMPT_ASSEMBLER=0) after the live chat smoke. Tests pin it
|
||||
// OFF so the broad chat-route suite keeps exercising the stable legacy
|
||||
|
||||
Reference in New Issue
Block a user