Explicit Semantic Generative Retrieval: A New Paradigm for E‑Commerce Search

Background

Search advertising and e‑commerce retrieval have moved from term matching to semantic modeling. Traditional inverted indexes are limited to literal matches, while dense‑vector retrieval lacks interpretability for fine‑grained ad control. Generative retrieval that forces models to emit opaque Semantic IDs (SIDs) forces large language models (LLMs) to memorize artificial codes, wasting pre‑training knowledge and degrading zero‑shot generalization.

Instead, the system builds a stable, controllable, and explainable attribute‑based semantic space on the product side and aligns queries to this space, enabling structured‑attribute recall.

Technical solution: asymmetric single‑sided alignment

After fixing the product‑side semantic space, the core problem is high‑quality query alignment. Prior work such as GRAM assumes symmetric information density between query and product, which fails because products contain rich multimodal data (title, OCR, detail, SKU) while queries are short (e.g., red Dress).

Research NanoVDR (Mar 2026) shows that symmetric joint training leads to information collapse in such asymmetric scenarios. The solution adopts an asymmetric paradigm: the product side acts as a fixed semantic anchor (Teacher) and the query side learns a one‑way projection (Student).

Engineering layer (output constraints) : model outputs must belong to a predefined candidate set, guaranteeing exact match with the inverted index.

Business layer (behavior distillation) : click and exposure data guide the query model to distill user intent into product attributes.

This “fixed semantic space + one‑way projection” avoids information collapse, reduces system complexity, and lowers online inference cost.

Product‑side multimodal feature extraction

A two‑stage pipeline “Multi‑Source Independent Extraction + LLM‑as‑Judge” is used.

Four independent sources (Title, OCR, Detail, SKU) are sent separately to the LLM to preserve modality‑specific signals.

Extraction prompts enforce three constraints:

Structured base constraint : ~20 high‑frequency attributes (brand, product_type, material, function, etc.) form the extraction baseline.

Category‑aware reasoning : the LLM must use the product’s category path for logical inference (e.g., under “Power Bank”, “140W” becomes function: 140W Fast Charging).

Open‑world long‑tail discovery : an additional_attributes array captures attributes beyond the base schema.

After independent extraction, a second‑stage LLM “Quality Control Specialist” merges, de‑duplicates, and normalizes values. Example: raw ["Lc11, Lc11"] is corrected to ["LC11"].

Schema construction & candidate value standardization

Raw KV pairs are noisy. A category‑constrained schema is built for AliExpress’s ~30 top‑level categories using an LLM multi‑agent cross‑validation mechanism:

Statistical purification : count coverage and top‑value distribution per key.

Multi‑model debate : different LLMs propose core target_key and assign importance tiers.

Cross‑evaluation & human arbitration : engineers resolve disagreements to produce a global schema.

Value standardization follows a five‑step “hard‑rule first + LLM‑assisted” pipeline:

Data profiling to detect JSON fragments, separators, marketing prefixes, or numeric noise.

Custom cleaning & normalization per high‑priority attribute.

Loss interception: dropped high‑PV values are reviewed by the LLM and, if valid, trigger code patches.

Coverage analysis & decision: split attributes into soft (vector‑absorbed) and hard (exact‑match) groups.

Production: generate canonical_value dictionaries and index mappings.

Query‑side intent structuring & dynamic candidate alignment

With the teacher space fixed, short queries are translated into the standard dictionary via three stages:

Semantic structuring : a category‑aware LLM parser (Qwen‑Max) extracts both generic and specific terms and enforces hierarchy extraction. Example: product_category: ["Phone", "Xiaomi Phone"] and brand: "Xiaomi".

Dynamic candidate alignment : first, BERT vector search retrieves the top‑5 raw‑value candidates; then the LLM judges equivalence and relevance, discarding mismatched items.

Behavioral distillation : historical click logs are merged into the teacher data, providing an implicit “behavior vote” that complements LLM semantics.

Student model training: 14B SFT + DPO

Training a 14B LLM as the Student model proceeds in two stages.

SFT (Supervised Fine‑Tuning) : high‑confidence strict samples form the base; loose samples are patched with click‑distribution signals (entropy ≤ 0.5, Wilson > 0.15).

DPO (Direct Preference Optimization) : positive samples are strict KV pairs present in the candidate pool; negative samples are SFT outputs that fall outside the pool. DPO penalizes “out‑of‑space” generations, forcing the model to obey the constrained dictionary.

Lexical sorting of multi‑value attributes guarantees a unique target token sequence.

Online recall & engineering deployment

Inference follows a “heavy‑offline, light‑online” philosophy:

Full‑recall (OR logic) : each attribute value triggers an unrestricted inverted‑index pull.

Match‑count ranking : items matching more query attributes rank higher.

Category filtering : predicted category acts as a safety filter to prevent cross‑category drift.

Structural miss‑recall (e.g., product labeled mini dress while query expects dress) is mitigated by offline expansion of lower‑level tags to their upper‑level counterparts.

Experimental results

Offline hit‑rate : native 72B extraction 15.3 %; native 72B + schema regularization 27.19 %; 14B SFT + DPO 42.1 %.

Online A/B test : revenue‑per‑mille (RPM) shows a clear uplift over the baseline bucket.

Conclusions

Grounding generative retrieval in an explicit, product‑side semantic space yields interpretable, controllable, and high‑quality recall. Key insights are:

Prioritizing semantic‑space governance over model complexity.

Asymmetric single‑sided alignment outperforms symmetric co‑training in highly imbalanced multimodal scenarios.

Generative capabilities must be tightly constrained through candidate pools, high‑confidence SFT data, and DPO alignment.