Skip to main content
Okapi BM25

Okapi BM25

Search complete. 28 mentions across 14 episodes found for "Okapi BM25".

Nov 18, 2026

James DooleyGUEST
45:34
But in local, it might work great.
James DooleyGUEST
45:36
But then you've got also in local, mainly it's query string matching or what's known as being BM25.
James DooleyGUEST
45:44
which is the quantitative algorithms, which means it's very simple.
James DooleyGUEST
45:48
It's old school methodology.
James DooleyGUEST
45:50
It's doing things like correlation.
James DooleyGUEST
45:53
So if you use things like onpage.ai or marketmoves or surfer SEO, in these BM25 algorithms, they work great because you just get the right engrams, get the right entities, get the right attributes on a page, almost TFIDF.
James DooleyGUEST
46:11
old school, like keyword stuffing can still work in a quantitative algorithm.
James DooleyGUEST
46:18
But the minute you start getting traffic or it's in your money, your life, or it's in a highly competitive market and it moves in and burr and mum and these more expensive algorithms start to come and check to see what your page is about.
Jeff WangGUEST
2:46
How does it differ from a database, just a regular database, like a SQL database or something? Yeah, I mean, I
Ryan DonovanHOST
2:51
know for a lot of old search, you'd use keyword algorithms like BM25 that process the text and had things like that.
Ryan DonovanHOST
2:58
But for a regular database, you search on a sort of unique index ID, right?
Jeff WangGUEST
3:04
Yeah, that's right.

Unknown podcast

Azure AI Search Hybrid Retrieval - Simply Explained

Nov 12 · 4 Mentions

Mirko PetersHOST
3:58
Once you type a query, the system scores how well each document matches.
Mirko PetersHOST
4:02
That scoring method has a name, BM25, and you can understand it without touching the math.
Mirko PetersHOST
4:07
BM25 gives more weight to words that appear often inside one document, and less weight to words that appear in almost every document.
Mirko PetersHOST
4:14
Take the word THE for a second.
Mirko PetersHOST
4:15
It shows up nearly everywhere, so it tells you almost nothing.
Mirko PetersHOST
4:18
A rare code like XC4471 shows up once, so it tells you a great deal.
Mirko PetersHOST
4:22
BM25 captures that instinct automatically, so keyword search stays strong on part numbers, exact names, dates, and specialized terms.
Mirko PetersHOST
4:30
A small example shows how that plays out.
Connor ShortenHOST
41:59
So I think the next topic, maybe also talking about re-ranking, is agentic search.
Connor ShortenHOST
42:06
And yeah, we mentioned earlier kind of the success of Grep with agentic search, BM25, and the argument is kind of like, the agent is doing this iterative learning about the corpus while it's searching.
Connor ShortenHOST
42:16
And it gets this great feedback from Grepper BM 25 because it has like this no matches found.
Connor ShortenHOST
42:22
And there's a pretty cool paper that's called the questions gambit.

Unknown podcast

EP532: MemFit Fix Context Rot

Nov 4 · 1 Mention

speaker_0UNKNOWN
3:43
During the storage phase, Memfit indexes every single verbatim interaction using dual signals.
speaker_0UNKNOWN
3:48
It uses a lexical BM25 index to catch exact keywords or specific dates, and a semantic dense index to capture the underlying meaning.
speaker_0UNKNOWN
3:56
Now, it does do a tiny bit of background summarization to group episodes into segments, but crucially, those summaries are just pointers back to the raw verbatim text.
speaker_0UNKNOWN
4:06
The heavy, expensive LLM only gets involved at the very end of phase two, when you actually need it to answer a question.
Danqi ChenGUEST
82:43
And this model kind of worked really well And it can largely outperform the traditional IR retrieval models.
Danqi ChenGUEST
82:50
If you see this figure, so the blue curve here is a traditional IR approach, like a BM25 approach.
Danqi ChenGUEST
82:57
And the other curve, the orange curve, is basically training this kind of retrieval using only 1,000 question and answer pairs.
Danqi ChenGUEST
83:05
So by looking at all these different curves, basically using different number of training examples, so it's actually largely outperforming the traditional IR models.
Christopher PottsHOST
17:39
You'll see over here RAM requirements.
Christopher PottsHOST
17:41
So BM25, It has no hardware requirements at all.
Christopher PottsHOST
17:45
You can run that on anything.
Christopher PottsHOST
17:48
Whereas these models down here that have these really high MRR scores, hugely expensive in terms of compute.
Christopher PottsHOST
18:48
This is kind of interesting too.
Christopher PottsHOST
18:49
This is from the same paper.
Christopher PottsHOST
18:51
Here's BM25.
Christopher PottsHOST
18:53
costs essentially nothing, but it has very low performance.
Saahil JainGUEST
45:48
One is kind of your standard lexical indices.
Saahil JainGUEST
45:50
So these ones are indices that essentially are using signals like BM25, as well as other You know, you can think of them as like heuristics, you know, related to keywords and term frequencies, et cetera, that allow you to find the right documents.
Saahil JainGUEST
46:04
So this is generally referred to as lexical retrieval.
Saahil JainGUEST
46:07
And then there's something called semantic retrieval, which is based on embeddings.
Kyle RoofGUEST
64:17
And the reason I like that is because, you know, the guys that put Yandex together were Google engineers and they were trying to approximate Google and as close as they can for the least amount of money.
Kyle RoofGUEST
64:28
And we found out that they're using BM25.
Kyle RoofGUEST
64:31
And BM 25 is just TF IDF plus word count.
Kyle RoofGUEST
64:34
And I've been screaming that for years.
Jacob X. LiGUEST
16:27
Um, we kind of have to search iteratively, and we have to decide which one actually, like when we condition on the search results, which part of the search results is actually going to contribute to the actually downstream research.
Jacob X. LiGUEST
16:41
So one finding from this mini experiment was that even if you use a very outdated model, the model, and given one of the worst retriever, BM25, Compared to a better model, like a newer model, these two models will pretty much cover the same set of papers.
Jacob X. LiGUEST
17:07
So basically, if I give it a title and some new content, the model will launch a lot of searches.
Jacob X. LiGUEST
17:14
And if you go through all the searches trajectories, both models will cover pretty much like a a set of paper that's believed to be useful.
Connor ShortenHOST
18:53
And then you're saying like, therefore it knows how to search more effectively.
Connor ShortenHOST
18:57
And I guess just that topic of like, how do you search more effectively? How do agents learn to search more effectively? What does it really mean to like write good search queries? I also find that topic to just be really interesting.
Connor ShortenHOST
19:09
And I guess the other one talking about like using BM25 as a search tool compared to say maybe vector search or late interaction models.
Connor ShortenHOST
19:20
I know we're maybe getting a little off topic, but just if you don't mind humoring me a little more, I love this kind of agentic search and just taking down this topic.

4 more episodes mention Okapi BM25.

Create an account to see the whole feed, search across every transcript, and follow the entities you care about.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.