A multi-intent modelling framework for search
Reframing ranking from a single black-box model into composable intent models — the FY23–FY25 vision.
- +20%search CSAT
- Multi-intentframework
- Text-first3-year vision
Takealot Search handled the lion's share of customer journeys into the catalogue — and the team owned the result quality every one of those customers experienced. The technical surface was Elasticsearch; the harder problem was framing.
Every ranking conversation devolved into 'just make it better'. There was no shared vocabulary between product, engineering, and data science for what 'better' meant. And the query mix was wildly heterogeneous — customers searching for a category, a brand model, a gendered category, a colour, or a combination — yet a single ranking model was expected to serve them all.
Decompose. Treat search ranking not as one model but as a layered framework of independent intent models. Each intent gets its own model, its own data, and its own lifecycle.
- Set the 3-year scope (FY23–FY25): focus the framework on text search. Voice and visual deferred pending competitor + customer + technical research. VR / AR explicitly out of scope.
- Define the vision as a multi-intent modelling framework: the customer query passes through parallel intent models (book/non-book, category, brand, colour, gender, …). Each model resolves its own signal; results merge downstream.
- Worked example used as the shared vocabulary: 'black timberland boots for men' resolves to 'non-book', 'Boots', 'Timberland', 'black', 'male'. Stakeholders could now reason about a query at the intent level.
- Each intent is independent — added, removed, or retrained on its own. This unlocks parallel team workstreams and makes ranking improvements composable rather than monolithic.
- Anchored prioritisation in 'Impact Rank' analysis: query-type distribution + add-to-cart rate by type, so the team picked intents with the highest expected lift first rather than the most novel ones.
Multi-intent modelling framework — vision diagram (after slide 6).
+20% search CSAT over the period. The bigger structural win was a shared language — every ranking conversation now started with 'which intent are we tuning?' instead of 'just make it better'.
Framework made it possible to ship intent models incrementally and run intent-level A/B tests, which de-risked rollouts. The same architecture later absorbed semantic-retrieval work without re-platforming.
The biggest lesson wasn't technical. Making the *structure* of ranking legible to non-data-scientists is what unlocked stakeholder alignment. Without that, the data-science work would have shipped slower and been harder to defend.