2
0
Cortex + dbt = No More Manual Minutes: Ever wondered how to automatically transform your company’s messy meeting notes into crisp, actionable minutes—and deliver them right to your inbox, without a single spreadsheet or copy-paste? In this blog, we showcase a proof-of-concept (POC) that transforms unstructured meeting notes into insightful summaries using only Snowflake Cortex, dbt, and Snowpark—with no external APIs involved.
What We Built
Our architecture is AI-native, fully within the Snowflake ecosystem:
-
Data Cleaning with dbt
We start with raw meeting notes—realistic, noisy, and unstructured. Using dbt models, we clean the data, flag important cues like the presence of action items or key decisions, and prepare it for enrichment. -
Enrichment via Snowflake Cortex (LLM)
A single SQL function call within our dbt model leverages Snowflake Cortex COMPLETE to:-
Summarize long-form notes
-
Extract action items
-
Reformat the content into readable, professional minutes
-
-
Daily Digest with Snowpark Python
Using Snowpark, we schedule and send out a daily email digest of summarized meetings—fully automated and cloud-native.
Why is this unique?
-
End-to-end in Snowflake: No need to leave the warehouse. Everything—from transformation to LLM enrichment to email automation—is handled within Snowflake using native tools.
-
AI-native dbt: We designed dbt models that call LLMs directly using SQL. This allows seamless integration of AI into your ELT workflows.
For this demo: We synthetically generated 100+ meeting notes in Snowflake using SQL to simulate realistic meeting scenarios—covering daily standups, sprint planning, and stakeholder reviews
Technical Implementation:
Technical Implementation:
- dbt Staging Model—Transform and Flag
Purpose:
- Standardize and clean incoming meeting notes
- Flag records likely to contain action items
2. dbt LLM Model—Summarize and Extract With Cortex
Purpose:
- Use Snowflake Cortex’s LLMs (like llama2-70b-chat) to generate:
- 2–3 bullet-point meeting minutes
- Action items (free text and structured)
Once the model gets executed we see MEETING_MINUTES_LLM table in snowflake;
NOTES_RAW(captured from meeting)
Quarterly planning: Rahul presented OKRs. Nina to draft communications plan.
MEETING_MINUTES:(Generated using LLM)
Sure! Here are 2-3 bullet points summarizing the meeting notes you provided, focusing on decisions and key discussions:
* Quarterly planning: Rahul presented OKRs (Objectives and Key Results) for the upcoming quarter.
Communications plan: Nina was assigned to draft a communications plan to support the OKRs.
Next steps: The team agreed to review and finalize the communications plan at the next meeting.
ACTION_ITEMS:(Generated using LLM)
Sure! Here are the action items from the meeting note you provided:
* Nina: Draft communications plan
No other action items were mentioned in the meeting note. Therefore, the list of action items is:
* Nina: Draft communications plan
I hope this helps! Let me know if you have any further questions.
3. Snowpark Automation—Daily Email Digest
Purpose:
- Send a professional HTML email digest of all meeting summaries for the day, fully automated.
Meeting Mail:
By orchestrating dbt’s transformation power, Snowflake Cortex’s native LLM capabilities, and Snowpark’s seamless Python automation, we’ve built a modern analytics pipeline that goes far beyond dashboards or batch reports.
