Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Update: Source Download – Inspect, grab assets, screenshot and record video/GIF(chromewebstore.google.com)
    discuss
  2. Omens of Exceptional Talent(guzey.com)
    discuss
  3. Debian Inference Portal (Unofficial)(debian.net)
    discuss
  4. Understanding Runtime Governance for Enterprise AI Deployment
    discuss
  5. Investigate top humanity questions with your own AI agent(interworky.com)
    1comments
  6. Can open-source prompt-injection detectors catch realistic AI agent attacks?(github.com/rudratoshs)
    discuss
  7. AI Passport Photo(aipassportphoto.org)
    1comments
  8. How to Audit an AI Agent:Time-Travel Debugging and Drift Measurement(medium.com/mirarshadtalpur)
    discuss
  9. The cost and quality frontier across 29 AI coding models(bito.ai)
    discuss
  10. Inferact's Kimi K3 Megakernel hits 700 tok/s on TPU, Beating GPU(inferact.ai)
    1comments
  11. Europe's fastest supercomputer is finally getting its domestic silicon infusion(theregister.com)
    discuss
  12. FBI investigating hacking group's claim of breach of agent info(nbcnews.com)
    discuss
  13. Dolphin Progress Release 2609(dolphin-emu.org)
    discuss
  14. Show HN: I asked 5 LLMs 10k questions and indexed every domain they cited(uvansa.com)
    discuss
  15. Meta ditches the camera on its newest smart glasses(theverge.com)
    discuss
  16. The E-Ink Monopoly [video](youtube.com)
    discuss
  17. Blink – Embeddable one-pass typed decision in 33 µs(github.com/sqliteai)
    discuss
  18. What Changed from GPT-3.5 to GPT-4? GPT-3.5: 0/30 Empty Nulls. GPT-4: 30/30(reddit.com)
    discuss
  19. Steve Jobs' Full iPhone 4 Antennagate Press Conference Q&A Session(youtube.com)
    discuss
  20. Can You Beat Jev?(antics.gg)
    discuss
  21. Albanese says OpenAI hacked Medicare and told Australia months later(theguardian.com)
    discuss
  22. What's Left to Get Wrong(blogsh.app)
    discuss
  23. Meta is making a standalone Muse AI gadget(theverge.com)
    discuss
  24. Mysteries of AI Generalization(astralcodexten.com)
    discuss
  25. Lysios – LLM red teaming org(lysios.ai)
    discuss
  26. The biggest surprises from a year working at OpenAI(aalokpatwa.substack.com)
    discuss
  27. Ways to Encrypt Data on Servers(lwn.net)
    discuss
  28. C-to-Rust Fallacy: Automatic Refactoring != Memory Security(arxiv.org)
    discuss
  29. The Nightmare of Voskhod 2(smithsonianmag.com)
    discuss
  30. Show HN: FireWand – A lightweight manager for GCP SQL databases(firewand.dev)
    discuss

Semantics Delivery Network: Rethinking Web Retrieval for LLM Agents

1 pointsby 49m agoarxiv.org
1 comments
49m agoHN ↗

When using an AI agent to look up information, we usually just dump an entire webpage, website, or directory into the model and let it "brute force" its way through to pick out a few potentially useful paragraphs.

This paper proposes a different approach:

1. Use a query retrieval system to push only the relevant paragraphs to the AI. 2. Consolidate and deduplicate the information from those sections.

This approach cuts costs and runs much faster.

This method solves the efficiency problem, but accuracy cannot be guaranteed. It relies entirely on the capability of the query system mentioned in the first point here.

Another way to improve data accuracy is to extract and transform the semantic layer before data entry. This does make writing data slower, but semantic data is different from transactional business data.

Semantic data doesn't necessarily need to be consumed by operational systems immediately. From a database perspective, it doesn't require OLTP characteristics. Therefore, a better approach might be to let this data undergo a period of processing and cleaning to become meaningful semantic information before making it available for subsequent retrieval.