💬 Reach out if you want to know more about our study and related work at amass.
28.02.2025 09:06
👍 0
🔁 0
💬 0
📌 0
💬 Reach out if you want to know more about our study and related work at amass.
⁉️ What this means for your RAG applications:
1) Advanced re-rankers aren't always worth the cost.
2) Real-world data brings unique challenges not captured in standard benchmarks.
3) Domain- and dataset-specific evaluation is critical to decide when to use re-rankers or simpler alternatives.
📊 We found that computationally expensive neural re-rankers can struggle to outperform simple keyword matching in real-world scenarios. This is an issue because re-ranking is often seen as a way to "fix" impresice document search and retrieval.
🔍 Language model Re-rankers: Not Always Smarter Than Simple Keyword Matching!
Happy to share new applied AI research results from our team at amass in collaboration with great partners at Chalmers, University of Gothenburg, LMU Munich, and MCML.
Link: arxiv.org/abs/2502.17036
Kind reminder to sign up for our meetup on Tue next week to talk bits and bio in a great setting with awesome people. Please remember to also sign up on the Danish Industry website (linked to from the event description). Looking forward to seeing many of you!
...to decrease latency for each request, you need a bit more. Read more here and take this primarily as a nudge to run your own analyses for your application: www.alexanderjunge.net/blog/fastapi...
Basic async performance testing with FastAPI and Locust
I wrote a short, technical blog post comparing different approaches to implement async in Python FastAPI applications. The key takeaways is that sprinkling in async/await keywords improves throughput but...
Event: Copenhagen @bitsinbio.bsky.social meetup with talks from Reshape Biotech & Novonesis with a view of the Copenhagen skyline. Read more about the talks and event here (and remember to sign up to register via the DI website!): lu.ma/xf8jfd1e
I spent a bit of time digging into the prompting going on behind the scenes in PydanticAI, a new AI agent framework for Python released last week. Here is what I found: www.alexanderjunge.net/blog/show-me...
Recommend doing the same for any LLM tools you are considering.