9 karma · joined January 26, 2023
So I did. This is the full 6-week focus plan I would’ve wanted in my first B2B SaaS marketing job. It's heavy on frameworks, light on fluff, and painfully honest about where we lost time by doing too much too soon.
Figured it might help others in similar spots.
Entrepreneurs often shy away from data-heavy products because they believe they can't make it without the data mountains OpenAI or others have. I don't think that's true. I think, you can:
1. Create value with minimal data, by pulling in data you might not yet own (like external data sets, publically available ones, etc.). Then, use that to attract a minimal set of first users, get their data, and go from there incrementally.
2. You can heavily incentivize early data contributions as Glassdoor does.
3. And you can bootstrap with alternative data sources.
Of course, there are a few battles you shouldn't fight, but I think there's way more potential for data-heavy startups out there than people realize.
It contains a complete implementation - with scraping context data from a docs website - chunking it, getting embeddings via the openAI API - loading it into pinecone - and finally a simple Q&A interface with streamlit on top of it
Here's a quick summary: - LangChain and LlamaIndex are great tools for quick exploration - But aren't perfect for production-grade use - I think we all know the "LangChain is pointless" debate, but there's a lot of real meat to it, and Pat describes a few of them (a lot of LangChains extractors are super basic, 2-3 liners without retries etc.) - LLM applications are all about moving data, extracting and enriching data (creating embeddings!) are the most expensive ones of those steps - A bunch of data engineering tools are out there that make these two steps much easier, versionable, robust, and reproducible. - Meltano is one such tool and Pat implemented the above described pipeline with it
FWIW: The GitHub project that comes with the post is super easy to run and super modular. I just tested it and was able to modify everything for my own application within 30 mins.
Key insights:
ETL was super helpful a decade ago!
But the assumptions that made ELT super useful don't hold any more (expensive storage, limited compute abilities of databases)
Now the world is cluttered with lots of ETL marketing, making it hard to see what's really going on.
ELT should be the default choice, always (and the rise of dbt backs that up).
Companies still on ETL should consider a switch.
ELT and ETL are architecture patterns for data pipelines. ETL is usually implemented with one ETL or data integration tool; ELT is usually implemented with one EL and one T tool (e.g., dbt).
Beware of the words "ETL tool" or "ELT" tools. There aren't any good conventions yet on those. (While there are ETL tools that support the ETL pattern, ELT usually is built with at least two tools. So an ELT tool is either a T only or an EL only tool - more confusion!)
- backward induction (from game theory): if you got a million options on every step, start with the goal, and work backward. - first-principle thinking: data is a super young discipline, so there are no actual "best practices", go to the basics, and reduce your questions to things from other fields. - unlearning: Try to assume everyone is wrong. And then think about your problem again. - actions work without data: No, whatever you're building is not essential to anyone. People will always take action, even with crappy data. - datacisions cycle: data gets consumed and turned into action again in the "datacisions" cycle through information, insight & decision. Use the cycle to build your product, invest and understand trends. - .... many more...