@mrscoopers
Qdrant ๐ฅ | ๐ฅจ organizer of "Bavaria, Advancements in SEarch Development" (BASED) Meetup | love NLP & always end up working on smth search tech-related | latent PHD | certified yapper ๐
Qdrant joins the AI-Powered Search course as a guest lecturer!
Our @mrscoopers.bsky.social will speak on โMixing Sparse & Dense Representationsโ in Week 2 of Modern Retrieval for Humans and Agents by Trey Grainger and Doug Turnbull.
๐ Details๏ฟผ: aipoweredsearch.com/articles/ai-...
After seeing 3 questions full of confusion on
@qdrant.bsky.social Discord, I finally got a hint๐ซก
Threw together a guide on decay functions for #relevance score boosting, aka fresher by time/closer by geo
๐ชhttps://medium.com/qdrant/untangling-relevance-score-boosting-in-vector-search-ccffb4a574d0
Starting with:
โ๏ธone dense high-dim;
โ๏ธone dense small-dim;
โ๏ธone sparse neural;
โ๏ธone sparse lexical (BM25, obviously);
โ๏ธone multimodal (text + images).
๐ Demo youtu.be/nJIX0zhrBL4?...
Inferencing embeddings & using them in @qdrant.bsky.social in one go.
๐ฉ๏ธ #Cloud #inference dropped RN (local with FastEmbed was sneakily there for some time).
So here comes the convenient plugging in of text/images in Qdrant, while saving on latency.
๐A lb of details
The week of 22โ26 == the search week in #Berlin
Qdrant will host a full day of workshops & talks on #vector #search & #AI at Colosseum Berlin on the 26th.
Btw, itโs a Filmtheater, so weโll all be kinda movie stars (pun heavily intended).
๐ Details: qdrant.tech/blog/vector-...
Yesss, fin amazing talk on BASED Meetup turned into a cute blog!
P.S. Mostly got a stroke spelling "hitchhiker"
Will spill tea around @qdrant.bsky.social dreams to make Qdrant MCP server the default solution against coding assistant hallucinations๐
Next Wed, #Munich, lu.ma/xh0natks
Yooo, Sparse Neural Retrieval fi-na-lly getting the attention it deserves!
Kudos to @hf.co for making this 5th version of SentenceTransformers real, I really-really want to see what the HF community will craft with this ๐๐๐
๐ฅ huggingface.co/blog/train-s...
Attending Office Hours of @DataTalksClub's "LLM Zoomcamp" reminded me I'm not even in the vector search bubble, I am in a vector search snow globe...
I do be unironically loving people asking questionsโค๏ธ
P.S. A link to @LukawskiKacper & my "Vector Search" module JFYI ๐
๐ฅ
#bbuzz went beyond expectations (#MICES also were amazing)
Huge thanks to our hosts Tacto, the vibe was perfect!
Made me sad & jealous I donโt have such an office to go to๐
I hope we'll continue being THAT lucky & aiming for the next meetup in SeptemberโOctober!
๐ lu.ma/based_meetup
We had two amazing speakers:
๐ ๏ธ Clelia Astra Bertelli blew minds with her open-sourced AI engineer experience, condensed into 35 mins
๐ฉโ๐ Monica Riedler showed how multimodality in RAG really matters, backed by her research.
All materials are here:
๐ lnkd.in/dGnkwpqA
๐ฅจ/๐งต Yesterday, @wrigley_dan & I nailed the 2nd ๐๐๐๐๐ #search meetup in #Munich!
๐ It's heartwarming to recognize people walking in. Iโll proudly assume we have regulars already :)
At the same time, every new member is super precious:3
#RAG #AI #tech
๐video with hybrid search example www.youtube.com/watch?v=sYP_...
๐node itself www.npmjs.com/package/n8n-...
๐gitHub for starring github.com/qdrant/n8n-n...
Last Friday Qdrant dropped an official n8n node with e-v-e-r-y-t-h-i-n-g.
Was possible due to new n8n's amazing feature - verified community nodes.
No more HTTP nodes for calling Qdrant's API ever again!
@qdrant.bsky.social + @n8n.io = ๐ค
๐links
โค๏ธ I am sincerely excited, I love train rides, meeting my colleagues irl (remote work is slowly driving me nuts) & this conf!
P.S. @qdrant.bsky.social added a bit to the #BBuzz organization, so you'll see us on coffee breaks; our Kacper ลukawski will be taking care of showing what we've got:)
โฐ ๐๐ง๐ ๐ฐ๐๐๐ค ๐ญ๐ข๐ฅ๐ฅ ๐๐๐ซ๐ฅ๐ข๐ง ๐๐ฎ๐ณ๐ณ๐ฐ๐จ๐ซ๐๐ฌ
I'll really try to make my talk (on the 17th at 11:10 am in Kesselhaus) fit this "๐๐ฎ๐ฎ๐ฎ๐ฎ, ๐ ๐ด๐ฉ๐ฐ๐ถ๐ญ๐ฅ ๐ต๐ณ๐บ ๐ต๐ฉ๐ช๐ด ๐ณ๐ฏ" vibe of @berlinbuzzwords.de!
To TLDR, it's about making ๐ฌ๐ฉ๐๐ซ๐ฌ๐ ๐ฉ๐๐ซ๐ญ ๐ข๐ง ๐ก๐ฒ๐๐ซ๐ข๐ ๐ฌ๐๐๐ซ๐๐ก๐๐ฌ as ๐ซ๐๐ฅ๐ข๐๐๐ฅ๐ as BM25, but ๐ฌ๐๐ฆ๐๐ง๐ญ๐ข๐๐๐ฅ๐ฅ๐ฒ ๐๐ฐ๐๐ซ๐ (so ๐๐๐ญ๐ญ๐๐ซ than BM25)
<Doing video editing>
CapCut:
- Lemme quickly transcribe your "FastEmbed" as "fast & bad", ha?
- How about we call "upsert operation" an "absurd operation"?
Is CapCut secretly paid by our competitors?๐ค
Was at Stiven Wilson's concert this weekend
Here's a riddle for you: if an LLM finishes generation when it sees EOS, when do prog rock artists understand that the song should end?
Finished Thought-Action-Observation
- "In the case of LLMs fine-tuned for function-calling, the thought process is optional"
> the thought process is optional.
... me every other day.
Seriously, are we teaching agents to think before they act, when 90% of us here don't do that?
Finished tools & special messages & found my sweet spot of self-motivation: dumb questions into the void๐ค
1. Is there a good read on jailbreaking agents? (spooky-scary)
2. Are instruct models slowly transforming to instruct-to-call-tools ones?
#hf_agents_course #agents
Finished "What are LLMs" (bach & masters in ML certainly are a cheat code);
๐ค2 q's:
1. How ADHD are agents? How many steps do they need to forget what they are doing in the kitchen?
2. Do we already have a united prompting protocol? Does MCP solve it?
#hf_agents_course #agents
Will check out @huggingface.bsky.social's "Agents Course"
๐ซฃI think I've never finished any online course ever & am interested to see:
a) as a devrel, how ppl out there do courses;
b) as a vector search advocate, how ppl out there approach agents.
The 1st post here & about the stuff I am psyched about๐
๐ฅจAnother onsite event of my baby, BASED #search meetup, is confirmed & happening on 12th of June in Tacto's office
I love this community & us discussing search without gatekeeping & boring advertisements:)
๐ lu.ma/cyelt1n6
#ai #rag #agents
