Skip to main content

Flipkart Commerce Cloud Search

Product Documentation v1.0 Built By Retailers, For Retailers


Confidential — For Publisher and Partner Use Only © 2026 Flipkart Commerce Cloud


About This Document

This document is the product reference for Flipkart Commerce Cloud (FCC) Search. It covers the platform's architecture across four pillars — Organize, Understand, Rank, and Experience — the semantic search layer, capability status, strategic vision, and integration guidance for publishers and their technical teams.

Audience: Network publishers, e-commerce product managers, merchandising teams, and integration engineers.

Prerequisites: Familiarity with e-commerce concepts (product catalogs, search results pages, conversion metrics) is helpful but not required.


Table of Contents

  1. Introduction
  2. Platform Architecture — The Four Pillars
  3. Pillar 1 — Organize: Indexing and Taxonomy
  4. Pillar 2 — Understand: Query Processing Pipeline
  5. Pillar 3 — Rank
  6. Pillar 4 — Experience
  7. Semantic Search
  8. Capability Summary
  9. Market Context and Strategic Vision
  10. Roadmap
  11. Metrics and Measurement
  12. Onboarding Checklist
  13. Glossary

1. Introduction

Overview

FCC Search is a composable, multi-tenant e-commerce search platform built on the same engine that powers Flipkart — one of the world's largest commerce platforms by search volume. It is purpose-built for mid-to-large retailers and marketplaces seeking higher conversion on existing traffic, without the complexity and cost of building or maintaining a search infrastructure in-house.

Search is not a feature — it is the primary discovery mechanism for e-commerce. Roughly 40–55% of all units sold in e-commerce flow through organic search. Customers who search convert significantly more than those who browse, and their average order value is higher. A customer who hits a dead end — null results, wrong results, confusing results — does not try again. They leave.

FCC Search is designed to ensure that does not happen.

The Retail Store Analogy

The simplest way to understand what a search system does is to think of a large physical retail store. Every element of the store is designed to help a shopper find what they came for — products grouped by type, arranged on shelves in a logical order, placed in sections that match how shoppers think. When a shopper asks a staff member for help, they understand what was meant — even if the wrong word was used — and guide the shopper to the right aisle. Once in the aisle, the most popular or best-value products are at eye level, not buried at the back.

E-commerce search does all of this digitally:

  • Organize — How the store is laid out and how products are shelved
  • Understand — The staff member who interprets your query and points you in the right direction
  • Rank — Which products are placed at eye level vs. buried on the bottom shelf
  • Experience — What information is visible on the shelf tag to help you decide

FCC Search is built around these four pillars.

Who This Platform Is For

PersonaDescription
Network PublishersRetailers and marketplaces ($100M–$2B GMV, 10K–5M SKUs) seeking higher conversion from existing traffic through best-in-class search relevance
Merchandising and Category TeamsBusiness users who need control over what products surface for which queries — without raising engineering tickets for every change
Integration EngineersTechnical teams integrating FCC Search APIs into storefronts, mobile apps, and analytics pipelines
E-Commerce Product ManagersTeams responsible for search quality, null search reduction, and conversion optimisation across discovery surfaces

2. Platform Architecture — The Four Pillars

Overview

Every FCC Search query passes through four sequential layers. Each layer has a distinct responsibility, and together they produce a ranked, relevant result page for every user interaction.

User Types a Query


┌─────────────────────────────┐
│ PILLAR 1 — ORGANIZE │ Products indexed into a demand-side store tree
│ (Indexing & Taxonomy) │ Real-time + batch ingestion pipelines
└──────────┬──────────────────┘


┌─────────────────────────────┐
│ PILLAR 2 — UNDERSTAND │ Sequential query processing pipeline
│ (Query Processing) │ DNA → Spell → Classify → Intent → Retrieve
└──────────┬──────────────────┘
│ Candidate Set

┌─────────────────────────────┐
│ PILLAR 3 — RANK │ 3-stage cascaded ranking (L0 → L1 → L2)
│ (Ranking) │ Coarse sort → product reranking → personalisation
└──────────┬──────────────────┘
│ Top N results

┌─────────────────────────────┐
│ PILLAR 4 — EXPERIENCE │ AutoSuggest, Snippets, Spotlights, Filters
│ (Search Results Page) │ Null handling, sort, grid/list view
└──────────┬──────────────────┘


Search Results Page (SRLP) — rendered by publisher storefront

Important: FCC Search is a headless, API-first service. It returns ranked result data via API — the publisher's storefront is responsible for rendering. This gives publishers complete control over the visual experience while FCC handles the intelligence layer.


3. Pillar 1 — Organize: Indexing and Taxonomy

Overview

Before any query can be answered, products must be indexed in a way that mirrors how shoppers think — not just how sellers list. The Organize layer structures the catalog into a demand-side taxonomy and makes every product retrievable in near real-time.

Status: ✅ Live

3.1 Store Tree (Demand-Side Taxonomy)

FCC Search organizes products into a Store Tree — a hierarchical taxonomy from L0 (All Products) down to Leaf Stores (specific product types). Every product maps to a Leaf Store, and the Leaf Store a query is classified into determines the entire universe of results returned.

All Products (L0)
├── Electronics (L1)
│ ├── Phones & Tablets (L2)
│ │ └── Smart TVs (L3 ● Leaf)
│ └── Computing (L2)
│ └── Laptops (L3 ● Leaf)
├── Home & Furniture (L1)
│ └── ...
└── Fashion (L1)
├── Men's T-Shirts (L3 ● Leaf)
└── Dresses (L3 ● Leaf)

Every product maps to a Leaf Store. The Leaf Store determines:

  • Which products are in the result universe for a given query
  • Which filters are available (Electronics shows RAM/storage; Fashion shows size/fit)
  • Which ranking signals are applied
  • How the Search Results Page (SRLP) is laid out

A wrong taxonomy classification means wrong results — always. Taxonomy quality is a direct multiplier on search quality. A product filed in the wrong store will never surface for queries that belong to the correct store.

Demand indexing vs. supply indexing. Products are indexed into stores based on how buyers think, not just the seller's vertical. A unisex t-shirt can live simultaneously in Men's, Women's, and Kids' stores. This demand-side mapping is configured by catalog and merchandising teams and is one of the highest-leverage levers in the system.

3.2 Indexing Pipeline

FCC Search maintains two ingestion paths for the product index:

PathLatencyPurpose
Real-Time IngestionNear-real-timePrice changes, stock updates, availability flags. Staleness here is a direct revenue risk — a product shown as in-stock that is actually out of stock undermines buyer trust immediately.
Batch IngestionPeriodic (scheduled)Product attribute updates, taxonomy reclassification, image changes, new catalog additions.

Catalog quality directly impacts search quality. Sparse product attributes — missing titles, incomplete specifications, absent category assignments — make products effectively invisible to relevant queries even if they are correctly indexed. Publishers should prioritise catalog completeness as a foundational dependency for search performance.


4. Pillar 2 — Understand: Query Processing Pipeline

Overview

The Query Processing Pipeline fires sequentially on every query. It transforms raw user input into a structured retrieval instruction that the index can act on. The order of steps is critical — an error in an early stage propagates through all subsequent stages and degrades the final result set.

Raw Query


[DNA] → [Spell Correction] → [Store Classification] → [Intent Understanding]


[Query Rewrite] → [Quantifier Extraction] → [Partial Match]


[Lexical Retrieval + Semantic Retrieval] → Candidate Set for Ranking

4.1 DNA — Do Not Augment Gate

Status: ✅ Live

The first step in the pipeline is a protection gate. DNA identifies brand names, trademarks, product codes, and other terms that must not be modified by any downstream module. Brand names that clear DNA are passed through unchanged — they are never spell-corrected, rewritten, or expanded.

This protection is essential: spell-correcting a brand name like "Defy" or "Cemento" into a common dictionary word corrupts everything that follows.

4.2 Spell Correction

Status: ✅ Live (with ongoing improvements)

Spell Correction fixes genuine typos in queries that cleared the DNA gate. The system distinguishes between a real misspelling and an unusual-looking brand name — getting this wrong corrupts all downstream processing.

FCC uses a combination of a trained spell correction model augmented with brand-specific override lists to protect known brand names from incorrect correction. New brand terms are onboarded to the override list to maintain protection as the client's catalog evolves.

4.3 Store Classification

Status: ✅ Live

Store Classification is an ML model that predicts which Leaf Store a query belongs to. This is the single most consequential step in the pipeline — the classified store determines the entire result universe.

InputOutput
Corrected user queryPredicted Leaf Store (e.g., "washing machine" → Washing Machines Leaf)

When the model's confidence for any Leaf Store falls below a threshold, the query falls back to the parent-level store — a broader result universe with lower precision. This fallback protects against null results at the cost of some relevance precision.

FCC's Store Classifier is trained on client-specific query and catalog data. For new client onboardings, the classifier is retrained on the publisher's local query patterns, taxonomy, and brand ecosystem — a key step in the onboarding process.

4.4 Intent Understanding

Status: ✅ Live

Intent Understanding classifies the semantic type of the query. Different query types receive different retrieval and ranking treatment:

Intent TypeExampleRetrieval Treatment
Brand"Nike"Brand-filtered retrieval; brand store prioritised
Category"running shoes"Broad category retrieval within classified store
Feature"earphones under 2k"Feature + price constraint extraction
Thematic"gym essentials"Broad semantic matching across categories

Understanding intent allows the system to optimise both what is retrieved and how results are ranked and presented for each query type.

4.5 Query Rewrite / Query Expansion

Status: 🔜 Coming Soon

Query Rewrite expands search recall without sacrificing precision. It adds semantically equivalent terms to the retrieval query — for example, "hiking footwear" triggers retrieval also including "trekking shoes" and "trail boots."

This capability is critical for tail queries where lexical matching alone fails. Without it, queries using non-standard terminology for a product category return no or poor results even when matching products exist in the catalog.

4.6 Quantifier Extraction (Numeric Constraint Handling)

Status: 🔜 Coming Soon

The Quantifier module extracts numeric constraints from natural-language queries and encodes them into the retrieval query. Examples:

  • "TVs under 30,000" → price ceiling constraint applied at retrieval
  • "laptops 40k–60k" → price band constraint applied at retrieval
  • "65-inch TV" → screen size constraint applied at retrieval

Without this capability, numeric terms in queries are treated as raw text and ignored — the price or size constraint the user expressed is not applied to results.

4.7 Partial Match

Status: 🔜 Coming Soon

Partial Match is a null-rescue mechanism. If a query containing multiple tokens returns no results after all processing steps, Partial Match progressively relaxes the query — dropping less essential tokens and loosening constraints — until a result set is found.

This directly reduces null search rate. Without it, multi-term queries that partially match the catalog return zero results even when relevant products exist.

4.8 Retrieval — Lexical and Semantic

Status: ✅ Live

FCC Search uses two parallel retrieval signals that fire simultaneously and whose outputs are merged before ranking:

SignalHow It WorksStrength
Lexical RetrievalToken-match query against the product index. Matches on title, attributes, brand, and category terms.High precision for exact and near-exact matches. Fast.
Semantic RetrievalDense vector match by meaning using the FK V5 embedding model (LLM-augmented). Products are matched by semantic intent, not just token overlap.Captures intent for natural-language, long-tail, and synonym-heavy queries that lexical retrieval misses.

Both signals contribute to the candidate set passed to ranking. See Section 7 for a dedicated description of the Semantic Search layer and its current rollout status.


5. Pillar 3 — Rank

Overview

Ranking determines which products in the candidate set are shown at the top of the results page. Position 1 receives approximately 40% of clicks; position 10 receives approximately 2%. Ranking decisions have an outsized impact on both conversion and revenue outcomes.

FCC Search uses a three-stage cascaded ranking architecture, where each stage narrows candidates with progressively richer signals.

~1M Candidate Documents


┌────────────────────────────────────────────────┐
│ L0 — Coarse Sort │ ✅ Live
│ Primary signal: Units Per Impression (UPI) │
│ Fast, broad filter. Output: ~10K candidates │
└──────────────────────┬─────────────────────────┘


┌────────────────────────────────────────────────┐
│ L1 — Collapse Sort + Product Reranking │ ✅ Live
│ Buy box selection (1 best listing per product)│
│ Speed + Quality + Price + Seller signals │
│ Output: top 1,000 products │
└──────────────────────┬─────────────────────────┘


┌────────────────────────────────────────────────┐
│ L2 — Personalisation │ 🔜 Coming Soon
│ Near-real-time session + purchase history │
│ Top 120 results shown on SRLP │
└────────────────────────────────────────────────┘

5.1 L0 — Coarse Sort

Status: ✅ Live

L0 applies a fast, broad scoring to the full candidate set retrieved from the index. The primary signal is UPI (Units Per Impression) — a measure of how frequently products in a position have historically converted to purchases. L0 eliminates clearly irrelevant or low-quality candidates before the more expensive L1 signals are applied.

5.2 L1 — Collapse Sort and Product Reranking

Status: ✅ Live

L1 performs two operations:

Buy Box Selection: For any product sold by multiple sellers, L1 selects the single best listing to represent the product in results — the "buy box" winner. This collapses multiple seller listings of the same product into one result card.

Product Reranking: The collapsed product list is re-scored using a richer signal set:

Ranking SignalDescription
UPI / fRPIUnits Per Impression with time-decay. Penalises products whose historical popularity is stale; gives recently popular products a fair weighting.
SpeedDelivery speed to the user's location. Faster delivery receives a ranking boost — a key commercial signal for platforms where delivery SLA is a differentiator.
Listing Quality Score (LQS)Completeness of the product listing: title richness, image count, attribute fill rate. Incomplete listings are demoted.
Product Quality Score (PQS)Review count and average rating. High-quality, well-reviewed products are surfaced above equivalent products with sparse or poor reviews.
Price Value Score (PVS)Competitive pricing relative to the category median. Products priced at or below the category median receive a ranking boost.
Seller TierSeller's historical fulfilment reliability, return rate, and responsiveness. Higher-tier sellers receive a ranking advantage.

Configurable signal weights: L1 signal weights are configurable per publisher — different weight profiles can be applied for different business contexts (e.g., sale events, category promotions). A self-serve ranking configuration layer is currently in development to allow business users to adjust these weights without engineering intervention.

5.3 L2 — Personalisation

Status: 🔜 Coming Soon

L2 applies personalisation signals to the top 120 products from L1 to produce the final ranking shown on the SRLP. It uses near-real-time signals from the user's current session and historical purchase behaviour to re-order results for that specific user.

L2 personalisation is expected to deliver meaningful uplift for returning users and high-frequency shoppers, where session context strongly predicts purchase intent.

5.4 Offline Ranking Calibration

Status: 🔜 Coming Soon

FCC is building an offline simulation system that runs ICU (Impressions, Clicks, Units) simulations to identify optimal signal weight configurations for L0/L1 ranking. This system will replace the current manual approach to signal weight setting with a principled, data-driven calibration process that can be run after every major catalog or traffic change.


6. Pillar 4 — Experience

Overview

The Experience layer covers everything the user sees — before, during, and after typing a query. A perfectly ranked result set that is hard to navigate, filter, or understand still fails. The Experience layer ensures the right information reaches the user in a way that makes choosing easy.

Status: ✅ Live (all components in this section)

6.1 AutoSuggest

AutoSuggest provides real-time query completions as the user types. Suggestions are drawn from a combination of popular queries, trending searches, past user queries, and category shortcuts.

FeatureDescription
Popular completionsHighest-volume query completions for the typed prefix
Trending suggestionsQueries gaining momentum in the last 24–72 hours
Category shortcutsDirect links to relevant category stores based on the prefix
In-session contextSuggestions adapt based on what the user has already searched or browsed in the current session
Zero-prefix stateBefore any character is typed, the system surfaces personalised starting suggestions based on user history and platform trends

AutoSuggest has its own ranking model — it is not simply a UI feature but a ranked retrieval problem in its own right.

6.2 Sort and Filter

Sort options and faceted filters allow users to refine results after the initial search.

Sort options:

OptionDescription
RelevanceDefault — system's ranked order balancing all ranking signals
Price: Low to High / High to LowPrice-sorted results
PopularitySorted by UPI / purchase volume
RatingSorted by average review rating
Newest FirstRecently added products surfaced at the top

Filters: Filter sets are Store-specific and generated dynamically from the index. Electronics surfaces RAM, storage, screen size, and connectivity filters. Fashion surfaces size, color, fit, and material filters. The filter quality is directly dependent on catalog attribute completeness — products with sparse attributes do not contribute to filter facets and effectively become unfilterable.

6.3 Snippets

Snippets are structured information cards displayed on each product result — showing key attributes, price, delivery estimate, and rating without requiring the user to click through to the product page.

Snippets reduce click-through friction and help users make faster decisions at the results level. Snippet quality scales directly with catalog attribute richness: products with complete, structured attributes produce rich snippets; sparse products produce empty or minimal cards.

6.4 Spotlights

Spotlights are dynamic callout labels displayed on product cards that communicate urgency and value signals at a glance:

Spotlight TypeTrigger
Limited StockProduct inventory falls below a defined threshold
NewProduct was recently added to the catalog
More 4 LessProduct is part of a promotional or value offer

Spotlight signals are dynamic — they are computed in real time from inventory and promotions data rather than being manually applied labels.

6.5 Null Search Handling

Null search — queries that return zero results — is a top health metric for every search system. A user who hits a null result does not retry; they leave.

FCC's null search handling re-engages user intent rather than ending the session:

ComponentDescription
Related category suggestionsLinks to stores relevant to the failed query
Alternative query suggestions"Did you mean X?" suggestions derived from similar successful queries
Partial match fallback(Coming Soon) Progressive query relaxation to return near-matches when exact matches fail

Null search rate is measured and tracked continuously. The introduction of semantic search, spell correction improvements, and the planned Partial Match module each target null rate reduction as a primary outcome metric.

6.6 Grid and List View

The results page layout adapts to the Store context:

ViewDefault ContextRationale
GridFashion, Accessories, Home DécorVisual-first categories where product images are the primary decision driver
ListElectronics, Appliances, B2BDetail-first categories where specifications and price comparisons matter

Publishers can configure default view preferences per Store and allow users to toggle between views.


Overview

Semantic Search is FCC's most significant search quality investment — a dense vector retrieval layer that matches queries by meaning, not just by token overlap. It is currently live and being progressively rolled out to full traffic.

Status: ✅ Live — Rolling Rollout (30% → 100%)

How Semantic Search Works

Traditional lexical retrieval answers the question: "which products contain words that appear in the query?" Semantic retrieval answers a different question: "which products mean the same thing as what the user asked for?"

A user searching for "gym shoes" on a platform that lists products as "athletic footwear" or "training sneakers" gets no results from lexical retrieval — the words don't match. Semantic retrieval understands that these phrases describe the same thing and surfaces the right products.

ApproachRetrieval LogicStrengthWeakness
LexicalToken matching against indexHigh precision for exact matches. Fast.Fails on synonyms, natural language, long-tail queries
SemanticDense vector similarity by meaningCaptures intent; handles natural language and synonymsRequires model training; higher compute cost
Blended (Coming Soon)Both signals merged into a single ranked result setBest of both — precision + recallRequires careful merging and calibration

The FK V5 Embedding Model

FCC's semantic retrieval is powered by the FK V5 embedding model — a two-stage trained, LLM-augmented model built on Flipkart's corpus of 200M+ products and years of user interaction data. The model generates dense vector representations of both queries and products, enabling semantic similarity matching at scale.

Key properties of the V5 model:

  • Domain-specific training: Trained on e-commerce query and product data — not a general-purpose language model. Understanding of commerce-specific semantics (brands, categories, specifications) is built in.
  • LLM augmentation: A second training stage uses LLM-generated signals to improve concept-level matching for abstract and thematic queries.
  • Client adaptation: The model can be fine-tuned on client-specific catalog and query data during onboarding to improve performance for the publisher's category mix and market.

Current Rollout Status

Semantic search is currently live at 30% of traffic. The current implementation uses a position-split model — lexical retrieval fills positions 1–20 of the result set, and semantic retrieval contributes from position 21 onwards.

Next milestone — Blended Retrieval (Coming Soon): The position-split model is being replaced with true blended retrieval, where lexical and semantic signals are merged into a single unified ranking across all positions. This eliminates the artificial boundary and allows the best signal to win at every position in the result set.

What Semantic Search Improves

Query TypeWithout SemanticWith Semantic
Synonym queries"hiking footwear" returns no results if catalog uses "trekking shoes"Returns relevant products regardless of terminology used
Natural-language queries"something to charge my laptop fast" returns null or unrelated resultsMaps to fast chargers and USB-C adapters correctly
Long-tail queriesMulti-token descriptive queries frequently hit nullMeaning-based matching finds relevant products even for unusual phrasing
Thematic queries"home office setup" fails lexical matchingReturns monitors, chairs, desks, and accessories through semantic clustering

Primary outcome metric: Null search rate reduction. Secondary metrics: result quality uplift for long-tail queries (measured by position-level click data) and conversion rate improvement for semantic-served result positions.


8. Capability Summary

CapabilityPillarStatus
Store Tree (Demand-Side Taxonomy)Organize✅ Live
Real-Time Price / Stock IndexingOrganize✅ Live
Batch Attribute / Catalog IndexingOrganize✅ Live
DNA — Do Not Augment GateUnderstand✅ Live
Spell CorrectionUnderstand✅ Live
Store Classification (ML Model)Understand✅ Live
Intent UnderstandingUnderstand✅ Live
Lexical RetrievalUnderstand✅ Live
Semantic Retrieval (V5 Model)Understand✅ Live — Rolling (30% → 100%)
Query Rewrite / Query ExpansionUnderstand🔜 Coming Soon
Quantifier Extraction (Numeric Constraints)Understand🔜 Coming Soon
Partial Match (Null Rescue)Understand🔜 Coming Soon
Blended Retrieval (Lexical + Semantic)Understand🔜 Coming Soon
L0 Coarse SortRank✅ Live
L1 Collapse Sort + Product RerankingRank✅ Live
Configurable Ranking SignalsRank🔜 In Progress
Result Explainability (Client-Facing)Rank🔜 In Progress
L2 PersonalisationRank🔜 Coming Soon
Offline Ranking CalibrationRank🔜 Coming Soon
AutoSuggestExperience✅ Live
Sort and FilterExperience✅ Live
SnippetsExperience✅ Live
SpotlightsExperience✅ Live
Null Search HandlingExperience✅ Live
Grid / List ViewExperience✅ Live
Natural Language / Conversational SearchExperience🔜 Coming Soon
Audience Manager Signal IntegrationExperience🔜 Coming Soon
Multi-Tenant Sandbox (New Client Demo)Platform🔜 In Progress
Search Generalisation (Multi-Tenant)Platform🔜 In Progress

9. Market Context and Strategic Vision

The Opportunity

E-commerce search is a $8B global TAM, with a realistic addressable market of $800M across FCC's primary geographies (US, UAE, Singapore, ANZ). FCC's conservative 3-year SOM target is $45M across these four markets.

Geography3-Year SOM TargetKey Context
United States$26M~1,200 ICP accounts · largest mid-market depth globally
ANZ$9M~320 accounts · English-first · underpenetrated by specialists
UAE / GCC$7M~180 accounts · Hybris EoL tailwind · strong SaaS appetite
Singapore$3M~90 accounts · SEA hub · gateway to regional marketplaces

Ideal Customer Profile (ICP)

FCC Search is optimised for mid-market, business-led retailers and marketplaces:

AttributeProfile
Revenue scale$100M – $2B GMV
Traffic1M – 50M monthly sessions
Catalog size10K – 5M SKUs
Team3–15 person digital or merchandising team; small or no dedicated search engineering team
CategoriesT1 (Electronics, Auto, B2B MRO, Pharmacy) and T2 (Fashion, Beauty, Home, Grocery)
Buyer postureBusiness-led (CPO, Head of Digital, VP E-commerce) · wants conversion outcomes, not infrastructure knobs

What these clients want: Outcome-first dashboards, merchandising control without engineering tickets, fast onboarding, predictable pricing, and composable integration with their existing stack.

What FCC uniquely delivers: Baseline relevance from session one without requiring client data to cold-start; transparent merchandising controls; a free sandbox for evaluation before commitment; category-specific tuning; and agent-ready architecture.

Competitive Landscape

CompetitorProfileFCC Differentiation
AlgoliaTech-led SMB, ~$15–20K ACVFCC targets business-led buyers; Algolia's buyer is the developer. FCC wins on merchandising controls and commerce depth.
Constructor.ioMid-large business-led, ~$250–400K ACVConstructor is the lone serious specialist in FCC's segment. FCC undercuts on pricing and wins where Constructor has thin presence — UAE, Singapore, ANZ.
BloomreachMid-large suite, ~$150–300K ACVHeavy bundled product; slow deployments. FCC is composable — plug in what the client needs, not the whole suite.
SFCC / AdobeEnterprise suite, $500K+ ACVBundled search add-ons of uneven quality. FCC wins when clients are actively re-platforming or dissatisfied with their suite's search.
CoveoEnterprise B2B, ~$300–500K ACVEnterprise-only; legacy orientation. FCC wins in mid-market where Coveo doesn't play.

6 Strategic Bets

BetDescription
1 — Composable Stack + Baseline RelevanceOne reusable search engine across clients; usable relevance from session one without client data cold-start
2 — Business GlassboxMerchandising rules, A/B testing, revenue attribution, and result explainability — from black-box to client-legible
3 — Frictionless Sales MotionFrom paid 6-week MVP to free 24-hour sandbox — cut pilot-to-close time dramatically
4 — Agentic Commerce ReadinessConversational on-site search surface + MCP/AP2 gateway for off-platform AI agents. Defend the agent-era pipeline.
5 — Market Tuning for US / UAE / SG / ANZRegion-specific language models, category packs, and compliance — beat Constructor where they barely play
6 — Winning Categories vs. Generalised SearchDeep category packs (electronics, fashion, horizontals) — depth beats breadth in mid-market

Core keyword search share will shrink. Three forces are reshaping discovery:

  • On-platform AI discovery: Shoppers shifting from "type and scroll" to "ask and get answered." Amazon Rufus, Walmart Sparky, and Flipkart Flippi are early signals of a broad shift toward conversational on-site search.
  • Off-platform AI discovery: Discovery starting in ChatGPT, Perplexity, and Gemini rather than on the retailer's site. 500M+ ChatGPT weekly active users are a pool of buyers whose first product interaction may never touch a retailer's search bar.
  • Agentic commerce: Software agents that research, compare, and transact without a browser. Gartner projects ~25% of shopping happening via agents by 2028.

FCC Search's agentic commerce readiness work (Bet 4) positions clients to capture discovery across all three surfaces — not just traditional on-site search.


10. Roadmap

Q2 2026 — Active Priorities

BucketItemStatus
Semantic SearchFull rollout to 100% traffic✅ Live → Rolling
Semantic SearchBlended retrieval (replace position-split model)🔜 Planned
RankingConfigurable ranking signals (self-serve for clients)🔄 In Progress
RankingOffline → online ranking plan improvements🔜 Planned
NL SearchNatural language / conversational query support🔜 Planned
PersonalisationAudience Manager signal integration🔜 Planned
Client TransparencyClient-facing ranking and result explainability🔄 In Progress
Sales EnablementSandbox environment for new client evaluation🔄 In Progress

Longer Arc — Search Generalisation

Beyond Q2, the strategic north star is making FCC Search a genuinely portable, multi-tenant product. This means abstracting every client-specific assumption — brand lists, store taxonomy, query patterns, ranking signal weights — into configurable, tenant-aware layers. A new client in a different country, with a different catalog and brand ecosystem, should be onboardable without rebuilding the system from scratch.

Search Generalisation is the foundational work that unlocks new client acquisition at scale.


11. Metrics and Measurement

Core Search Health Metrics

MetricDefinitionTarget Direction
Null Search Rate% of queries returning zero resultsMinimise
Query VolumeTotal search queries per day / per sessionTrack; volume drop signals UX or platform issue
Click-Through Rate (CTR)% of search sessions where a user clicks a resultMaximise
Conversion Rate (CVR)% of search sessions that result in a purchasePrimary business outcome metric
Average Order Value (AOV)Average revenue per order from search-originated sessionsTrack for uplift from personalisation features

Relevance Metrics

MetricDefinition
NDCG (Normalised Discounted Cumulative Gain)Ranking quality metric — measures how well the ranked result list matches an ideal ordering based on relevance judgements
Revenue per SearchRevenue attributed to search sessions ÷ number of search sessions
Position-Level CTRClick-through rate at each result position (P1, P2, ... P10). Position 1 should receive ~40% of clicks in a well-tuned system.
Semantic vs. Lexical Attribution% of clicks and conversions coming from semantic-served vs. lexical-served result positions

Null Search Diagnostic Metrics

MetricDefinition
Brand Query Null RateNull rate specifically for queries classified as brand-intent
Category Query Null RateNull rate for queries classified as category-intent
Tail Query Null RateNull rate for queries with low historical search volume (long-tail)
Store Classifier Fallback Rate% of queries falling back to parent store due to low classification confidence

Measurement Approach

Baseline definition. Before any new feature is rolled out, a measurement baseline — null rate, CTR, CVR — must be established. FCC works with publishers to define this baseline during onboarding.

Staged rollout. New ranking and retrieval features are rolled out in traffic tranches (e.g., 10% → 30% → 100%) with client sign-off at each stage. Metrics are reviewed at each gate before the next tranche is enabled.

Query-level analysis. For suspected quality issues, FCC provides query-level debugging access showing the full processing pipeline output for any specific query — which store was classified, which products were retrieved, and how they were ranked.


12. Onboarding Checklist

Overview

Standard FCC Search onboarding takes approximately 4 to 8 weeks from contract finalisation to first live traffic, depending on catalog size, taxonomy complexity, and how much training data is available for model adaptation.

Phase 1: Taxonomy and Index Setup

TaskOwner
Define Store Tree: L0 → L3 Leaf Stores for the publisher's category structurePublisher Catalog Team + FCC
Map demand-side stores to supply-side verticals (which stores are products indexed into?)Publisher Catalog Team + FCC
Configure filter sets per Leaf Store (which attributes to expose as facets)Publisher Catalog Team + FCC
Set up real-time indexing pipeline for price and stock updatesPublisher Engineering
Set up batch indexing pipeline for catalog attributes and taxonomy changesPublisher Engineering
Validate catalog attribute completeness for key categories (impacts snippet and filter quality)Publisher Catalog Team + FCC

Phase 2: Model Adaptation and Query Processing

TaskOwner
Retrain Store Classifier on publisher's taxonomy and local query patternsFCC Data Science
Configure DNA (brand/trademark) override lists for publisher's brand ecosystemPublisher + FCC
Build spell correction model and initial brand override listFCC Data Science + Publisher
Configure intent understanding for publisher's category and brand mixFCC
Adapt semantic retrieval model (V5) for publisher's catalog distributionFCC Data Science
Validate query processing pipeline end-to-end with sample query setFCC + Publisher

Phase 3: Ranking Configuration

TaskOwner
Define ranking signal weights for the publisher's business context (speed, price, quality priorities)Publisher + FCC
Configure Spotlight triggers (inventory thresholds, promotional signals)Publisher + FCC
Configure sort options and default sort per StorePublisher + FCC
Establish null search rate baseline before go-liveFCC + Publisher
Define measurement framework: which metrics, attribution windows, reporting cadencePublisher + FCC

Phase 4: Go Live and Optimisation

TaskOwner
Launch at partial traffic (10–30%) for initial quality validationPublisher + FCC
Monitor null search rate, CTR, and CVR for the first 2 weeksPublisher + FCC
Review position-level click data to assess ranking qualityFCC
Expand semantic search rollout based on validated metricsFCC + Publisher
Plan phased rollout of Coming Soon capabilities (Query Rewrite, Partial Match, L2 Personalisation)Publisher + FCC
Establish quarterly search health review cadencePublisher + FCC

Glossary

TermDefinition
AutoSuggestReal-time query completion suggestions displayed as the user types, drawn from popular queries, trending searches, and session context
Blended RetrievalA retrieval approach where lexical and semantic signals are merged into a single unified ranked result set, rather than being applied to separate position ranges. Coming Soon.
Buy BoxThe single best seller listing selected to represent a product in search results when multiple sellers offer the same item
Cascaded RankingA multi-stage ranking architecture where each stage applies progressively richer signals to a progressively smaller candidate set (L0 → L1 → L2)
CTRClick-Through Rate — the percentage of search result impressions that result in a user clicking a product
CVRConversion Rate — the percentage of search sessions that result in a purchase
DNADo Not Augment — the first gate in the query processing pipeline. Protects brand names, trademarks, and product codes from modification by downstream modules.
FCCFlipkart Commerce Cloud
FK V5 ModelFCC's domain-specific, LLM-augmented semantic embedding model trained on Flipkart's e-commerce corpus. Powers semantic retrieval.
Headless SearchFCC's deployment model — a search API that returns ranked result data; the publisher's storefront controls rendering and display.
ICPIdeal Customer Profile — mid-market ($100M–$2B GMV) business-led retailers and marketplaces in T1/T2 categories across US, UAE, Singapore, and ANZ
Intent UnderstandingThe pipeline step that classifies a query's semantic type — brand, category, feature, or thematic — to determine appropriate retrieval and ranking treatment
L0 / L1 / L2The three stages of cascaded ranking: L0 (Coarse Sort), L1 (Collapse Sort + Product Reranking), L2 (Personalisation)
Leaf StoreThe lowest-level node in the Store Tree taxonomy. Every product maps to a Leaf Store. The classified Leaf Store determines the result universe, available filters, and SRLP layout for a query.
Lexical RetrievalToken-based matching of query terms against the product index. Precise for exact and near-exact matches.
MACHMicroservices, API-First, Cloud-Native SaaS, Headless — FCC Search's architectural standard
MCP GatewayModel Context Protocol gateway — an interface that exposes client catalog, inventory, and pricing data to AI agents (ChatGPT, Perplexity, Gemini) for off-platform discovery. Part of the agentic commerce roadmap.
NDCGNormalised Discounted Cumulative Gain — a standard ranking quality metric measuring how closely a ranked list matches an ideal relevance ordering
Null Search RateThe percentage of search queries that return zero results. A primary search health metric.
NymeriaFCC's planned offline simulation system for ranking calibration — runs ICU (Impressions, Clicks, Units) simulations to identify optimal signal weights. Coming Soon.
Partial MatchA null-rescue mechanism that progressively relaxes a query (drops tokens, loosens constraints) until a result set is found. Coming Soon.
PQSProduct Quality Score — a ranking signal based on product review count and average rating
PVSPrice Value Score — a ranking signal that boosts products priced at or below the category median
Quantifier ExtractionA pipeline step that extracts numeric constraints from queries (price limits, size specifications) and encodes them into the retrieval query. Coming Soon.
Query RewriteA pipeline step that expands queries with semantically equivalent terms to improve recall without losing precision. Coming Soon.
SandboxA demo-ready evaluation environment where prospective clients can plug in their catalog, run test queries, and see FCC Search results in real time — before any commercial commitment. In Progress.
Search GeneralisationThe platform work to make FCC Search a portable, multi-tenant product by abstracting all client-specific assumptions into configurable tenant layers. In Progress.
Semantic SearchDense vector retrieval that matches queries by meaning rather than token overlap. Powered by the FK V5 embedding model.
SnippetsStructured product information displayed on result cards (key attributes, price, delivery, rating) without requiring click-through to the product page
SpotlightsDynamic callout labels on product cards (Limited Stock, New, More 4 Less) computed in real time from inventory and promotions data
SRLPSearch Results Listing Page — the page rendered by the publisher's storefront showing search results
Store ClassifierAn ML model that predicts which Leaf Store a user's query belongs to. The most consequential single step in the query processing pipeline.
Store TreeThe demand-side taxonomy hierarchy (L0 → L3 Leaf) that organises the product index. Products are indexed into stores based on how buyers think, not just how sellers list.
UPIUnits Per Impression — the primary L0 and L1 ranking signal. Measures how frequently a product in a given position has historically converted to a purchase.

© 2026 Flipkart Commerce Cloud — Confidential. For Publisher and Partner Use Only.