2026-01-29
Hard numbers for the search pipeline
Disclaimer
I used to think search/discovery/recommendation algos only did embedding search. I have now realised there's an entire pipeline, and not a single step in the pipeline can be skipped.
- Start with the entire text internet
- Use hard-coded initial list of people/blogs/urls/keywords to filter
- Use embedding search to filter further
- Use inference with both smart prompts (like paul graham prompts, surprising difference prompts) and the user context, to filter further
- Show the post to N other users, to filter further
Hard numbers
- hard-coded list - hetzner charges $200/mo for 1 gbps 32 thread machine which can max out egress instead of CPU AFAIK => ($200 / (30 * 86400)) / (10^9 / 8/5) = $0.000003/1M tokens (pay per query). Also pay disk cost of hosting entire full-text DB (unless someone else is hosting it)
- embedding search - openai text-embedding-3-small costs $0.01/1M tokens (pay once). Also pay disk cost of hosting entire vector DB
- inference - openai gpt-5.2 high costs $1.75/1M tokens (pay per query)
- use actual test users - $10 CPM => $10/(1000 views * 100 tokens/view) = $100/1M tokens/view (pay per query)
The text internet contains atleast 500B tokens.
Tips
- Even if you work on some much smaller dataset (like just lesswrong or just substack or whatever), you will definitely need to implement actual test users.
- You will probably also want to implement atleast one of either inference or embedding search.
- It might be easier to build on top of bluesky custom feeds or similar, as compared to building your own 'like' functionality from scratch for the entire internet (including across across multiple social media platforms that don't want to cooperate with each other)
Subscribe
Enter email or phone number to subscribe. You will receive atmost one update per month
Comment
Enter comment