Agentic Search

For complicated queries, like “Knicks gear from Nike or Adidas, for kids, in blue,” we recommend agentic search. Toggle to agentic mode, and an LLM plans the search: it extracts deterministic filters, may split the request into OR/AND sub-searches, runs them, and merges the results into one ranked list. Latency is higher than default or keyword (it’s LLM-bound), so use it for shopping and assistant flows where result quality matters more than milliseconds.

import Channel3 from "@channel3/sdk";

const client = new Channel3(); // reads CHANNEL3_API_KEY from env

const results = await client.products.search({
  query: "Nike or Adidas hoodies in black or grey for men, size large",
  config: { mode: "agentic" },
});

from channel3_sdk import Channel3

client = Channel3()  # reads CHANNEL3_API_KEY from env

results = client.products.search(
    query="Nike or Adidas hoodies in black or grey for men, size large",
    config={"mode": "agentic"},
)

curl -X POST https://api.trychannel3.com/v1/search \
  -H "x-api-key: $CHANNEL3_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "Nike or Adidas hoodies in black or grey for men, size large", "config": {"mode": "agentic"}}'

When to use

Prefer agentic when the request has several moving parts that should become real filters:

Multi-constraint shopping language — brand, retailer, category, size, price, and attributes in one sentence
Alternatives and branches — "Nike or Adidas…", "green 4-wheel or black 2-wheel…"
Conversational / vibe phrasing — descriptive language stays in the semantic query while constraints become structured filters
Pasted identifiers — SKU, MPN, GTIN, UPC, or EAN that should pin and boost an exact match

If you already know exact attribute handles and values, build them yourself with attribute filters — agentic is for when the caller has natural language, not a pre-built filter tree.

What it extracts

Natural language is turned into structured SearchFilters (and related ranking signals), including:

Extracted as	Examples in the query
Product type → category	`"TV"` → television (taxonomy-backed, not only a keyword boost)
Brand + retailer	`"Levi's"`, `"from Best Buy"` → catalog brand / website IDs
Gender + age	`"men"`, `"boys"`, and similar combined cues
Condition	`new` / `used` / `refurbished`, including `"vintage"` / `"pre-owned"`
Price min / max	`"under $400"`, `"above $100 or below $90"` (as separate ranges when disjunctive)
Colors	Named shades → product-realistic hex matching (not string equality on `"navy"`)
Category attributes	`"waterproof"`, `"wide feet"`, `"size 10"`, `"55 inch"`, `"queen size"` against that category’s schema
Physical dimensions	`"30 inches wide and 72 inches tall"` → structured L/W/H/weight filters when axes are named
Exact product IDs	SKU / MPN / GTIN / UPC / EAN → deterministic boost and pin
Stock preference	In-stock by default; widens when the user says `"even if sold out"`

Query shapes it plans for (not just single-filter extraction):

Alternatives / DNF — "Nike or Adidas hoodies in black or grey" → multiple sub-searches, then merge
Correlated constraints — "green 4-wheel or black 2-wheel suitcase" keeps color tied to the other trait per branch
AND within one product — "brown and yellow sneakers" stays one search with both colors
Descriptive residue — "cozy oversized cable knit sweater for winter" keeps vibe language in the semantic query while filters stay clean

Attribute coverage depends on the category taxonomy. Agentic does not replace hand-built attribute filters when you already know the exact handles and values.

The filters you provide override the filters set by the agent.

Same API surface

Request and response shapes are identical to default mode. Explicit filters combine with what the planner extracts, and pagination via next_page_token works as usual.

Getting Started

Make a Search

Get Product Data

Conversational Agent

Checkout

Pricing, Reporting, and Commissions

Build

Integrate

FAQ

When to use

What it extracts

Same API surface

​When to use

​What it extracts

​Same API surface

When to use

What it extracts

Same API surface