Sales ops · Production

Building Reffy: a RAG-based agent for Databricks sales and marketing to discover 2,400+ customer stories

The problem

Databricks sales and marketing teams could not efficiently find the right customer story at the right time — thousands of references were scattered across YouTube, databricks.com, LinkedIn, Medium, and internal slides with no unified search, story quality was opaque, and discovery relied on tribal knowledge, causing high-value references to be overused and relevant newer stories to be missed.

Workflow diagram · grounded in source
1
Multi-source story collection
integration
“collecting the text of stories from all of our data sources: we use standard Python webscraping libraries to gather YouTube transcripts, LinkedIn/Medium articles, and all public customer stories on databricks.com. Using Google Apps scrip…”
2
AI quality scoring
ai_action
“we classify the text by applying a rigorous 31-point scoring system (developed by our Value team) to each story via AI Functions. We prompt Gemini 2.5 to judge overall story quality by identifying the business challenge, the solution, th…”
3
Metadata extraction and tagging
ai_action
“The prompt also extracts key metadata like country and industry, products used, competition, and quotes—and tags stories based on whether they are publicly sharable or internal only”
4
Vector Search indexing
integration
“we sync this 'Gold' table to a Databricks Vector Search index, with the summary column containing all of the essential information an LLM would need to match customer stories to queries”
5
User submits query
trigger
“if you ask a question, you'll get a carefully thought-out answer with sources, but if you just enter a few keywords, Reffy will return top results in less than two seconds”
6
Agent hybrid search and reranking
ai_action
“we define a tool-calling agent that can look up the most relevant customer references with hybrid keyword and semantic search. We also use the Databricks re-ranker for Vector Search to improve results from RAG”
7
Personalized response delivery
output
“enabling users to discover and analyze over 2,400 Databricks customer references, delivering personalized responses, cross-story analysis, quotes, and more”
8
Usage monitoring and gap analysis
feedback_loop
“we go a step further to apply another AI Function to summarize the inputs and responses into recent themes and gap analysis. We want to understand which customer stories are popular and where we might have gaps, and the logs we collect f…”
Reported outcome

Reffy was adopted by over 1,800 Databricks sales and marketing employees who ran upward of 7,500 queries in its first two months, delivering more relevant and consistent storytelling, faster campaign execution, and confident at-scale use of customer proof.

Reported metrics
Users in first two monthsover 1,800
Queries run in first two monthsupward of 7,500
Customer references indexedover 2,400
Keyword search response timeless than two seconds
Show all 7 reported metrics
users in first two monthsover 1,800
queries run in first two monthsupward of 7,500
customer references indexedover 2,400
keyword search response timeless than two seconds
storytelling consistencymore relevant and consistent storytelling
campaign execution speedfaster campaign execution
infrastructure costs vs GPUsaving on infrastructure costs compared to GPUs
Reported stack
ReffyLakeflow JobsUnity CatalogDelta LakeAI FunctionsGemini 2.5Databricks Vector SearchDSPyMLflowDatabricks Model ServingAgent FrameworkReactFastAPILakebaseAI/BI DashboardDatabricks AppsDatabricks Asset BundlesGitHub ActionsGoogle Apps scriptsGoogle Sheet
Source
https://www.databricks.com/blog/tribal-knowledge-instant-answers-building-reffy-databricks
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Reffy was adopted by over 1,800 Databricks sales and marketing employees who ran upward of 7,500 queries in its first two months, delivering more relevant and consistent storytelling, faster campaign execution, and co…

What tools did this team use?

Reffy, Lakeflow Jobs, Unity Catalog, Delta Lake, AI Functions, Gemini 2.5, Databricks Vector Search, DSPy, MLflow, Databricks Model Serving.

What results were reported?

Users in first two months: over 1,800; Queries run in first two months: upward of 7,500; Customer references indexed: over 2,400; Keyword search response time: less than two seconds (source-reported, not independently verified).

How is this sales ops AI workflow structured?

Multi-source story collection → AI quality scoring → Metadata extraction and tagging → Vector Search indexing → User submits query → Agent hybrid search and reranking → Personalized response delivery → Usage monitoring and gap analysis.

Related sales ops cases