Simon Willison’s Weblog

Roaming RAG – make the model find the answers (via) Neat new RAG technique (with a snappy name) from John Berryman:

The big idea of Roaming RAG is to craft a simple LLM application so that the LLM assistant is able to read a hierarchical outline of a document, and then rummage though the document (by opening sections) until it finds and answer to the question at hand. Since Roaming RAG directly navigates the text of the document, there is no need to set up retrieval infrastructure, and fewer moving parts means less things you can screw up!

John includes an example which works by collapsing a Markdown document down to just the headings, each with an instruction comment that says <!-- Section collapsed - expand with expand_section("9db61152") -->.

An expand_section() tool is then provided with the following tool description:

Expand a section of the markdown document to reveal its contents.

- Expand the most specific (lowest-level) relevant section first
- Multiple sections can be expanded in parallel
- You can expand any section regardless of parent section state (e.g. parent sections do not need to be expanded to view subsection content)

I've explored both vector search and full-text search RAG in the past, but this is the first convincing sounding technique I've seen that skips search entirely and instead leans into allowing the model to directly navigate large documents via their headings.

# 3 am / ai, prompt-engineering, generative-ai, llms, rag

DSQL Vignette: Reads and Compute. Marc Brooker is one of the engineers behind AWS's new Aurora DSQL horizontally scalable database. Here he shares all sorts of interesting details about how it works under the hood.

The system is built around the principle of separating storage from compute: storage uses S3, while compute runs in Firecracker:

Each transaction inside DSQL runs in a customized Postgres engine inside a Firecracker MicroVM, dedicated to your database. When you connect to DSQL, we make sure there are enough of these MicroVMs to serve your load, and scale up dynamically if needed. We add MicroVMs in the AZs and regions your connections are coming from, keeping your SQL query processor engine as close to your client as possible to optimize for latency.

We opted to use PostgreSQL here because of its pedigree, modularity, extensibility, and performance. We’re not using any of the storage or transaction processing parts of PostgreSQL, but are using the SQL engine, an adapted version of the planner and optimizer, and the client protocol implementation.

The system then provides strong repeatable-read transaction isolation using MVCC and EC2's high precision clocks, enabling reads "as of time X" including against nearby read replicas.

The storage layer supports index scans, which means the compute layer can push down some operations allowing it to load a subset of the rows it needs, reducing round-trips that are affected by speed-of-light latency.

The overall approach here is disaggregation: we’ve taken each of the critical components of an OLTP database and made it a dedicated service. Each of those services is independently horizontally scalable, most of them are shared-nothing, and each can make the design choices that is most optimal in its domain.

# 5:12 pm / architecture, aws, databases, ec2, postgresql, s3, scaling, firecracker

New Gemini model: gemini-exp-1206. Google's Jeff Dean:

Today’s the one year anniversary of our first Gemini model releases! And it’s never looked better.

Check out our newest release, Gemini-exp-1206, in Google AI Studio and the Gemini API!

I upgraded my llm-gemini plugin to support the new model and released it as version 0.6 - you can install or upgrade it like this:

llm install -U llm-gemini

Running my SVG pelican on a bicycle test prompt:

llm -m gemini-exp-1206 "Generate an SVG of a pelican riding a bicycle"

Provided this result, which is the best I've seen from any model:

Here's the full output - I enjoyed these two pieces of commentary from the model:

<polygon>: Shapes the distinctive pelican beak, with an added line for the lower mandible.
[...]
transform="translate(50, 30)": This attribute on the pelican's <g> tag moves the entire pelican group 50 units to the right and 30 units down, positioning it correctly on the bicycle.

The new model is also currently in top place on the Chatbot Arena.

Update: a delightful bonus, here's what I got from the follow-up prompt:

llm -c "now animate it"

Transcript here.

# 6:05 pm / google, releases, svg, ai, generative-ai, llms, llm, gemini, pelican-riding-a-bicycle

Meta AI release Llama 3.3. This new Llama-3.3-70B-Instruct model from Meta AI makes some bold claims:

This model delivers similar performance to Llama 3.1 405B with cost effective inference that’s feasible to run locally on common developer workstations.

I have 64GB of RAM in my M2 MacBook Pro, so I'm looking forward to trying a slightly quantized GGUF of this model to see if I can run it while still leaving some memory free for other applications.

Update: Ollama have a 43GB GGUF available now. And here's an MLX 8bit version and other MLX quantizations.

Llama 3.3 has 70B parameters, a 128,000 token context length and was trained to support English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

The model card says that the training data was "A new mix of publicly available online data" - 15 trillion tokens with a December 2023 cut-off.

They used "39.3M GPU hours of computation on H100-80GB (TDP of 700W) type hardware" which they calculate as 11,390 tons CO2eq. I believe that's equivalent to around 20 fully loaded passenger flights from New York to London (at ~550 tons per flight).

# 6:30 pm / ai, generative-ai, llama, edge-llms, llms, training-data, meta, ollama