It support · Production

Nubank builds RAG-based internal knowledge search on Slack to reduce support tickets

The problem

Nubank's ~9,000 employees were spending significant time navigating fragmented per-team Confluence documentation and frequently opened support tickets just to find the right information or identify which team owned a topic.

Workflow diagram · grounded in source
1
Employee query via Slack
trigger
“The solution is integrated on Slack, so that employees can make your queries easily.”
2
Automated KB indexing
ai_action
“it is automatically extracted textual information from every Confluence department page. Each document is divided into chunks, retaining metadata such as its id, title, url and department. These chunks are converted into embeddings – num…”
3
First retrieval search
ai_action
“When a user performs a query, the same LLM converts the query into an embedding. This embedding is then used to identify and retrieve the most relevant chunks from the KB.”
4
Department router classification
routing
“the solution determines which department holds the relevant information to provide the answer. This is achieved through a classification engine powered by a LLM using a Dynamic Few-Shot Classification approach. LLMs are few-shot learners…”
5
Department-scoped second search
ai_action
“Once the domain is identified, a new search is triggered. In this step only the documents of the identified department are queried.”
6
Personalized answer generation
output
“The retrieved chunks and the user query are fed into the LLM in order for it to generate a personalized answer for the user. References for the URLs of the retrieved documents fed to the LLM are also provided to the user”
7
Ticket escalation fallback
routing
“If they are not happy with the answer, they are redirected to the portal of the department to raise a ticket on that matter.”
Reported outcome

The tool was adopted by most Nubankers, reducing both the time employees spend finding information and the number of support tickets created; the router achieves 78% precision and 77% recall, and 74% of internal domain answers were labeled accurate by department owners.

Reported metrics
Router precision78%
Router recall77%
Internal domain answer accuracy74%
Time to find informationreducing the amount of time to find the information they need
Show all 5 reported metrics
router precision78%
router recall77%
internal domain answer accuracy74%
time to find informationreducing the amount of time to find the information they need
support tickets createdreducing the number of tickets created
Reported stack
LLMRAGSlackConfluence
Source
https://building.nubank.com/ai-solution-for-search/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The tool was adopted by most Nubankers, reducing both the time employees spend finding information and the number of support tickets created; the router achieves 78% precision and 77% recall, and 74% of internal domai…

What tools did this team use?

LLM, RAG, Slack, Confluence.

What results were reported?

Router precision: 78%; Router recall: 77%; Internal domain answer accuracy: 74%; Time to find information: reducing the amount of time to find the information they need (source-reported, not independently verified).

How is this it support AI workflow structured?

Employee query via Slack → Automated KB indexing → First retrieval search → Department router classification → Department-scoped second search → Personalized answer generation → Ticket escalation fallback.

Related it support cases