Introducing The Palm Framework (Postgres + LightRAG + MCP)

In the unrelenting hype cycle of AI, I don't often get excited. Yes, LLM models are getting more and more powerful, yes context windows are getting bigger, yes, 'thinking' models are appearing more and more clever.

What do human centric AI applications look like?

But LLM's alone can't be amazing and integrated, human centric solutions. To do this, we need LLM's to be able to do 3 key things

1) Access an up to date and curated corpus of information that is relevant to the solution e.g. a legal documents database or a set of newspaper articles

2) The ability to really understand this information in both a detailed and general way, to understand and store the relationships between entities at a high and low level. e.g. an expert in politics will understand the relationship between liberalism and communism (high level) and also that Anthony Albanese is the member for Grayndler (low level).

3) The ability to reach out to related systems to both retrieve data and perform actions on behalf of the user.

and if the solutions are to be accessible

4) Technology should be available both open source and cost effective so small innovators can start crushing old problem domains

If you understand the space, you will know that there have been solutions for all of these capabilities for some time...

Retrieval Augmented Generation (RAG) is a technology that enables chunking and contextual indexing of large document corpuses.

Knowledge Graphs (and the notable implementation, GraphRAG) are a way of representing the relationship between entities

Application Programming Interfaces (APIs) are a technology that makes it possible for code to talk to other software systems for retrieval and action.

Elements of the Framework

There are now available, a combination of open source tools and standards that enable everybody to build solutions with these capabilities...

PostgreSQL

PostgreSQL has been around for over 35 years. It is an open source relational database. But that's not all, extensions to PostgreSQL enable amazing capabilities including Geospatial queries, Vector and Graph capabilities.

LightRAG

LightRAG is a technology that takes RAG and supercharges it with the ability to create knowledge Graphs. While generally similar to it's predecessor (Microsoft's GraphRAG), it is much cheaper to run, is open source and supports incremental update of the knowledge graph as new data arrives.

MCP (Model Context Protocol)

MCP is an emergent standard (now supported by industry leader OpenAI) to enable LLM's to reach out to MCP enabled endpoints to both retrieve data and take actions through those endpoints. It's not revolutionary but the fact that it has now emerged as the standard is fertilising a rich ecosystem of capabilities both on the endpoint side and on the client side.

LightRAG now supports PostgreSQL for vector and graph storage and Locally hosted LLM's via Ollama

..... It is now possible to build a software solution that has all the capabilities above but be 100% open source, be updated incrementally and be cost effective for small players.

If your mind isn't boggling right now, have a coffee or go for a walk..... your mind should definitely be boggled

Let's walk through an example

How can you turn "I want to sit under a palm tree in Fiji' into airline bookings?

1) PostgreSQL - has been pre-loaded (by LightRAG) with a knowledge graph and embeddings related to the travel domain.

2) LightRAG

• Takes the text of the question and searches the knowledge graph (with the help of embeddings) for entities and relationships that relate to the question
• Passes the question and the entity and relationships as context to the LLM

Fiji -[is a ]- Destination

Flights to Fiji -[take you to a]- Destination Fiji

Flights to Fiji -[Are run by]- Fiji Airlines

Flights -[need to be booked using]- a booking system

3) LLM

• takes the question, entity and relationships and determines that in order to action the question, it needs to make a flight booking.
• This instance of the LLM has been enabled to connect to the airline booking system via Model Context Protocol (MCP) and requests flight times from that interface.
• Facilitates a conversation with the user about times and dates, filling in blanks not presented in the original question.
• Makes a booking via MCP and informs the user

4) MCP

• Facilitates conversation with the LLM, calling internal API's to retrieve flight information and making bookings

When can I have this?

All of these technologies are available now. The only real bottleneck is MCP adoption. MCP is an emergent standard (November 2025). Companies that already expose API's can easily extend these to create MCP endpoints.

I my opinion, we are seeing the new way to do user interface, the new way to do software systems and we will soon think it perfectly normal to conduct our life admin by having a conversation with an AI.