When building a RAG system from scratch, which of the following take the most time and resources?

Gathering the data/documents for the knowledge base to be loaded17%

Curating, annotating, cleaning and/or pre-processing the data (excl. chunking)33%

Optimizing the splitter / chunking of the knowledge base documents

Defining the optimal RAG system architecture (vector/graph DB, critique models...)33%

Integrating the RAG system with the rest of the GenAI application tech stack17%

Other (grateful if you could specify in the comments!)

6 PARTICIPANTS

623 views1 Upvote1 Comment

Sort by:

Group Director of Information Security in Bankinga year ago

1. Absolute clarity about business use cases, the RAG is intended to resolve that would decide the repos to be loaded post curation. Laying down the governance model first.

2. If the query falls back to OpenAI Model, how much will it cost per prompt and who will bear those costs?

3. Data loss prevention despite curation and cleaning. Example: "What's the salary of x or y in my department?" Such data should be protected by an AI DLP solution and act as a proxy behind RAG.

Content you might like

Does your diversity, inclusion and belonging plan include practices and programs (i.e. retention and training programs)?

Yes80%

No16%

Unsure2%

View Results

Any suggestions for metrics that ensure data analytics teams are meeting organizational needs? How are you tracking your team’s performance?

What is a "must have" meeting for you or your team? If you can please share why/what makes it so great - topic, frequency, etc. - I'm looking to brainstorm some fresh opportunities for my group.

I am looking for some benchmark data on supplier onboarding cycle time to see how we measure up. I would like to include the time from intake to risk assessment to supplier created and ready to transact in our ERP.

What capabilities do you have to control AI access to data sources at your organization?

data security posture management 29%

data loss prevention 53%

data access governance 42%

encryption 36%

privacy enhanced technology 33%

use of synthetic data 11%

None, not using AI 2%