Что нового

The biggest single release yet, spanning search quality, observability, personalization, security and recommendations.

Search quality lab. Golden query sets with graded judgments, offline evaluation scoring nDCG, MRR, MAP, precision and recall, run-to-run comparison, and a regression gate that flags a ranking change before you promote it — with a webhook fired when a run breaks your configured threshold.

Request tracing and replay. Every search can be run with debug=true to get a full trace: which rules fired, which experiment variant applied, which rewrite strategy rescued a zero-hit query, and a per-hit explain of why it matched. The query log records every search, and any past query can be replayed against the current configuration to diff what changed.

Personalization platform. Behavioural profiles track category, brand and product affinities with time decay, derive segments automatically, and feed personalized boosts directly into ranking — configurable per collection, with a simulator to preview the effect before turning it on.

Reranking and fusion. Optional cross-encoder and LLM reranking on top of the existing hybrid search, plus a fusion layer supporting score-based, rank-based and query-adaptive blending between lexical and vector retrieval.

Rules lifecycle. Search rules now version every change, support one-click rollback, a dry-run preview before saving, and scheduled campaigns with recurrence — daily, weekly, or a custom window — for seasonal merchandising that turns itself off.

Security. API keys can now be scoped to specific collections, fields, filters and IP ranges. Custom roles with per-permission control, OIDC/Google/GitHub SSO, and SCIM 2.0 provisioning.

Recommendations. 16 recommendation model types, including query-to-item, upsell, cross-sell, sequential and graph-based models, with configurable blending between behavioural, vector and graph signals and automatic cold-start fallback.

Autopilot and alerts. The platform now proposes synonym, pin and boost suggestions from query and click behaviour, and alert rules watch zero-result rate, error rate and latency for anomalies.

The AI assistant now builds and reasons over a knowledge graph of your content, not just a flat set of retrieved chunks.

Documents are processed into entities and relationships, deduplicated with embedding similarity so "Acme Corp" and "Acme Corporation" resolve to one node, then clustered into communities with multi-level Louvain detection. A global search summarizes across communities for broad questions; a local search traverses entity neighborhoods for specific ones; DRIFT search combines both — a primer answer plus a small number of targeted follow-up searches — for questions that need both breadth and depth.

Multi-query retrieval now fuses results from several reformulated variants of a question using reciprocal rank fusion, and the zero-hit broaden_query escalation is now a cheap, staged ladder instead of one big relaxation step. Recommendation hydration also picked up MMR diversification, so a set of recommended items is less likely to be five near-duplicates of the same product.

Experiments now compute Wilson score confidence intervals on conversion and click-through metrics, so a variant's lift is reported with a real statistical bound instead of a raw percentage that might be noise. Search rules gained banner injection — trigger a banner by query match or facet filter, with priority ordering when more than one could fire — alongside the existing pins, boosts, exclusions and redirects. Recommendation models now report their own analytics: impressions, clicks, click-through rate, add-to-cart and revenue, broken down per model.

The console picked up organization team management (invite, assign roles, remove members), self-service API key creation and revocation, and a search analytics view reading directly from the platform's own event store: zero-result rate over time, top queries, and click-through rate per collection. An audit log viewer now lets an admin see who changed what, and when.

Search rules can now be scoped to the caller's device, language and country, so a pin, boost or banner can apply only to mobile traffic, only in a specific market, or only for a given locale — read from request parameters at query time, no separate configuration surface needed. Synonym suggestions mined from zero-result and click behaviour now have one-click Accept/Reject actions in the console. Storefronts using the assistant widget can also hand a conversation off to a live human support agent without leaving the chat.

Search rules landed as a first-class feature: pin specific documents to exact positions in the result set, cap the maximum number of results from a given field value (brand, seller, category) so one source can't dominate a page, and schedule a rule to a start and end time so seasonal changes turn themselves off. API keys used in browser widgets also picked up per-key rate limiting, on top of the existing per-organization quota, so one leaked key can't exhaust an entire account's request budget.

This is where AACSearch's tenant API went live: a contract-driven proxy under /v1 with per-organization tenant isolation, API-key authentication, and event ingestion feeding recommendation models straight out of the gate. Billing shipped alongside it — a credit wallet with usage-based accounting, subscription plans wired to per-tenant quotas, and both Stripe and Robokassa as payment providers, so a Russian-market storefront and a US-market one are equally first-class. The console launched as a Next.js app for authentication, organizations, billing, and tenant quota management, and the API reference moved to a proper Scalar UI generated straight from the OpenAPI contract.

Three new ways to search landed together: voice search (transcribe with Workers AI, then search by the transcript), image search (upload an image, get a caption and embedding, then search by meaning), and the AI assistant — a retrieval-augmented chat that cites the documents it draws its answer from, with conversation memory and streaming responses. Search rules also picked up exclusion filters and a zero-hits fallback, so a query with no matches can fall back to a recommendation model instead of an empty page.