The first workshops and talks from the AI Engineer Summit are now up! Join the >20k viewers on YouTube, find clips on Twitter (we’re also clipping @latentspacepod), and chat with us on Discord!Text-to-SQL was one of the first applications of NLP. Thoughtspot offered “Ask your data questions” as their core differentiation compared to traditional dashboarding tools. In a way, they provide a much friendlier interface with your own structured (aka “tabular”, as in “SQL tables”) data, the same way that RLHF and Instruction Tuning helped turn the GPT-3 of 2020 into the ChatGPT of 2022.Today, natural language queries on your databases are a commodity. There are 4 different ChatGPT plugins that offer this, as well as a bunch of startups like one of our previous guests, Seek.ai. Perplexity originally started with a similar product in 2022: In March 2023 LangChain wrote a blog post on LLMs and SQL highlighting why they don’t consistently work:* “LLMs can write SQL, but they are often prone to making up tables, making up field”* “LLMs have some context window which limits the amount of text they can operate over”* “The SQL it writes may be incorrect for whatever reason, or it could be correct but just return an unexpected result.”For example, if you ask a model to “return all active users in the last 7 days” it might hallucinate a `is_active` column, join to an `activity` table that doesn’t exist, or potentially get the wrong date (especially in leap years!).We previously talked to Shreya Rajpal at Guardrails AI, which also supports Text2SQL enforcement. Their approach was to run the actual SQL against your database and then use the error messages to improve the query: Semantic Layers to the rescueCube is an open source semantic layer which recently integrated with LangChain to solve these issues in a different way. You can use YAML, Javascript, or Python to create definitions of different metrics, measures and dimensions for your data: Creating these metrics and passing them in the model context limits the possibility for errors as the model just needs to query the `active_users` view, and Cube will then expand that into the full SQL in a reliable way. The downside of this approach compared to the Guardrails one for example is that it requires more upfront work to define metrics, but on the other hand it leads to more reliable and predictable outputs. The promise of adding a great semantic layer to your LLM app is irresistible - you greatly minimize hallucinations, make much more token efficient prompts, and your data stays up to date without any retraining or re-indexing. However, there are also difficulties with implementing semantic layers well, so we were glad to go deep on the topic with Artem as one of the leading players in this space!Timestamps* [00:00:00] Introductions* [00:01:28] Statsbot and limitations of natural language processing in 2017* [00:04:27] Building Cube as the infrastructure for Statsbot* [00:08:01] Open sourcing Cube in 2019* [00:09:09] Explaining the concept of a semantic layer/Cube* [00:11:01] Using semantic layers to provide context for AI models working with tabular data* [00:14:47] Workflow of generating queries from natural language via semantic layer* [00:21:07] Using Cube to power customer-facing analytics and natural language interfaces* [00:22:38] Building data-driven AI applications and agents* [00:25:59] The future of the modern data stack* [00:29:43] Example use cases of Slack bots powered by Cube* [00:30:59] Using GPT models and limitations around math* [00:32:44] Tips for building data-driven AI apps* [00:35:20] Challenges around monetizing embedded analytics* [00:36:27] Lightning RoundTranscriptSwyx: Hey everyone, welcome to the Latent Space podcast. This is Swyx, writer, editor of Latent Space and founder of Smol.ai and Alessio, partner and CTO in residence at Decibel Partners. [00:00:15]Alessio: Hey everyone, and today we have Artem Keydunov on the podcast, co-founder of Cube. Hey Artem. [0