Personalised Search Autocomplete

A search box that completes the same prefix the same way for everyone is a lookup table. A useful one ranks completions by what this user is likely to type, which means learning from the searches people have actually made. In Aito v2 that is a prediction over the query log: the typed prefix limits the candidates, and everything else you know about the session is evidence.

The queries below run against the live v2 sandbox. Its contexts table is a grocery shop's session log: one row per page view, with the queryPhrase the shopper had typed and a user link to users. Probabilities in the responses are rounded.

Popular completions for a prefix

Without personalisation, a completion is the share of past searches that start with the prefix. Open queryPhrase as candidates with get and filter by the prefix:

{
  "from": "contexts",
  "where": { "queryPhrase": { "$startsWith": "m" } },
  "get": "queryPhrase",
  "orderBy": "$p",
  "select": ["$value", "$p"],
  "limit": 5
}

{ "total": 8, "hits": [
  { "$value": "milk", "$p": 0.218 },
  { "$value": "meat", "$p": 0.206 },
  { "$value": "macaroni", "$p": 0.200 },
  { "$value": "myllyn paras", "$p": 0.182 },
  { "$value": "maxi", "$p": 0.082 } ] }

Eight distinct searches in the log start with "m", and the ranking is their frequency.

Personalised completions

Add the user. With predict, a condition on the predicted field itself restricts the candidates (only phrases starting with "m" can be answered), while the user condition is evidence that re-weights them:

{
  "from": "contexts",
  "where": { "user": "larry", "queryPhrase": { "$startsWith": "m" } },
  "predict": "queryPhrase",
  "select": ["$value", "$p"],
  "limit": 5
}

{ "total": 8, "hits": [
  { "$value": "meat", "$p": 0.412 },
  { "$value": "milk", "$p": 0.153 },
  { "$value": "macaroni", "$p": 0.133 },
  { "$value": "myllyn paras", "$p": 0.132 },
  { "$value": "maxi", "$p": 0.066 } ] }

Larry has searched for "meat" twice, so it moves to the top. For "veronica" the same query puts "macaroni" first (0.335), ahead of "milk" and "meat". The other candidates keep a share of the probability, so a user with little history still gets sensible completions rather than an empty list.

The same query in SQL, where LIKE 'm%' is the prefix condition:

SELECT * FROM predictions('contexts', 'queryPhrase',
  given => 'user = ''larry'' AND queryPhrase LIKE ''m%''', k => 5)

Why this completion

Add $why to see what moved the ranking:

{
  "from": "contexts",
  "where": { "user": "larry", "queryPhrase": { "$startsWith": "m" } },
  "predict": "queryPhrase",
  "select": ["$value", "$p", "$why"],
  "limit": 1
}

For "meat" the factor tree has three parts: the base rate of "meat" over the whole log, a normaliser that rescales it against the other candidates, and a lift of 1.86 from user: "larry". That last factor is the personalisation, stated as a number you can show or log.

Suggestions before the first keystroke

An empty search box can already suggest something. Exclude the empty phrase (page views without a search) and let the user be the only evidence:

{
  "from": "contexts",
  "where": { "user": "larry", "queryPhrase": { "$not": "" } },
  "predict": "queryPhrase",
  "select": ["$value", "$p"],
  "limit": 5
}

For Larry the top suggestions are "pirkka", "bread", "banana", "vegetable" and "paper". These probabilities are small (the largest is 0.076), because 117 different phrases compete; show them as suggestions, not as a confident guess.

Why Aito for this

The log is the model. Every search written to contexts counts in the next request. There is no suggestion index to rebuild.
Any evidence, same query. The user is one condition. The weekday, the current basket or the page type are more conditions in the same where.
Explainable ranking. $why separates popularity from personalisation.

Related: Smart search · Query Reference · Inference · Playground

← All v2 use cases