1
0
Fork 0
mirror of https://github.com/maybe-finance/maybe.git synced 2025-07-23 15:19:38 +02:00
Maybe/app/models/provider/openai/auto_merchant_detector.rb
Zach Gollwitzer 297a695d0f
Transaction rules engine V1 (#1900)
* Domain model sketch

* Scaffold out rules domain

* Migrations

* Remove existing data enrichment for clean slate

* Sketch out business logic and basic tests

* Simplify rule scope building and action executions

* Get generator working again

* Basic implementation + tests

* Remove manual merchant management (rules will replace)

* Revert "Remove manual merchant management (rules will replace)"

This reverts commit 83dcbd9ff0.

* Family and Provider merchants model

* Fix brakeman warnings

* Fix notification loader

* Update notification position

* Add Rule action and condition registries

* Rule form with compound conditions and tests

* Split out notification types, add CTA type

* Rules form builder and Stimulus controller

* Clean up rule registry domain

* Clean up rules stimulus controller

* CTA message for rule when user changes transaction category

* Fix tests

* Lint updates

* Centralize notifications in Notifiable concern

* Implement category rule prompts with auto backoff and option to disable

* Fix layout bug caused by merge conflict

* Initialize rule with correct action for category CTA

* Add rule deletions, get rules working

* Complete dynamic rule form, split Stimulus controllers by resource

* Fix failing tests

* Change test password to avoid chromium conflicts

* Update integration tests

* Centralize all test password references

* Add re-apply rule action

* Rule confirm modal

* Run migrations

* Trigger rule notification after inline category updates

* Clean up rule styles

* Basic attribute locking for rules

* Apply attribute locks on user edits

* Log data enrichments, only apply rules to unlocked attributes

* Fix merge errors

* Additional merge conflict fixes

* Form UI improvements, ignore attribute locks on manual rule application

* Batch AI auto-categorization of transactions

* Auto merchant detection, ai enrichment in batches

* Fix Plaid merchant assignments

* Plaid category matching

* Cleanup 1

* Test cleanup

* Remove stale route

* Fix desktop chat UI issues

* Fix mobile nav styling issues
2025-04-18 11:39:58 -04:00

146 lines
4.5 KiB
Ruby

class Provider::Openai::AutoMerchantDetector
def initialize(client, transactions:, user_merchants:)
@client = client
@transactions = transactions
@user_merchants = user_merchants
end
def auto_detect_merchants
response = client.responses.create(parameters: {
model: "gpt-4.1-mini",
input: [ { role: "developer", content: developer_message } ],
text: {
format: {
type: "json_schema",
name: "auto_detect_personal_finance_merchants",
strict: true,
schema: json_schema
}
},
instructions: instructions
})
Rails.logger.info("Tokens used to auto-detect merchants: #{response.dig("usage").dig("total_tokens")}")
build_response(extract_categorizations(response))
end
private
attr_reader :client, :transactions, :user_merchants
AutoDetectedMerchant = Provider::LlmConcept::AutoDetectedMerchant
def build_response(categorizations)
categorizations.map do |categorization|
AutoDetectedMerchant.new(
transaction_id: categorization.dig("transaction_id"),
business_name: normalize_ai_value(categorization.dig("business_name")),
business_url: normalize_ai_value(categorization.dig("business_url")),
)
end
end
def normalize_ai_value(ai_value)
return nil if ai_value == "null"
ai_value
end
def extract_categorizations(response)
response_json = JSON.parse(response.dig("output")[0].dig("content")[0].dig("text"))
response_json.dig("merchants")
end
def json_schema
{
type: "object",
properties: {
merchants: {
type: "array",
description: "An array of auto-detected merchant businesses for each transaction",
items: {
type: "object",
properties: {
transaction_id: {
type: "string",
description: "The internal ID of the original transaction",
enum: transactions.map { |t| t[:id] }
},
business_name: {
type: [ "string", "null" ],
description: "The detected business name of the transaction, or `null` if uncertain"
},
business_url: {
type: [ "string", "null" ],
description: "The URL of the detected business, or `null` if uncertain"
}
},
required: [ "transaction_id", "business_name", "business_url" ],
additionalProperties: false
}
}
},
required: [ "merchants" ],
additionalProperties: false
}
end
def developer_message
<<~MESSAGE.strip_heredoc
Here are the user's available merchants in JSON format:
```json
#{user_merchants.to_json}
```
Use BOTH your knowledge AND the user-generated merchants to auto-detect the following transactions:
```json
#{transactions.to_json}
```
Return "null" if you are not 80%+ confident in your answer.
MESSAGE
end
def instructions
<<~INSTRUCTIONS.strip_heredoc
You are an assistant to a consumer personal finance app.
Closely follow ALL the rules below while auto-detecting business names and website URLs:
- Return 1 result per transaction
- Correlate each transaction by ID (transaction_id)
- Do not include the subdomain in the business_url (i.e. "amazon.com" not "www.amazon.com")
- User merchants are considered "manual" user-generated merchants and should only be used in 100% clear cases
- Be slightly pessimistic. We favor returning "null" over returning a false positive.
- NEVER return a name or URL for generic transaction names (e.g. "Paycheck", "Laundromat", "Grocery store", "Local diner")
Determining a value:
- First attempt to determine the name + URL from your knowledge of global businesses
- If no certain match, attempt to match one of the user-provided merchants
- If no match, return "null"
Example 1 (known business):
```
Transaction name: "Some Amazon purchases"
Result:
- business_name: "Amazon"
- business_url: "amazon.com"
```
Example 2 (generic business):
```
Transaction name: "local diner"
Result:
- business_name: null
- business_url: null
```
INSTRUCTIONS
end
end