A talk by Jo Kristian Bergum (Hornet.dev) arguing that BM25 — the ~30-year-old lexical ranking function named after the 25th in a series of scoring experiments — remains highly effective for agentic search because the user, not the function, has changed. LLM agents can formulate long, specific queries and issue many in a row, unlike human two-to-three-word searches. Evidence includes a deep research benchmark of 830 riddle-like questions over ~100,000 web documents showing high accuracy when answer-bearing documents are placed directly in context, indicating reasoning is not the bottleneck — retrieval and query formulation are. Bergum likens context windows to floppy discs (~350k tokens before quality degrades) and advocates dumping retrieved documents into a file-system workspace for the model to explore with grep and similar primitives. Timestamps cover the BM25 name origin, benchmark results, query formulation, BM25 variants, the file-system pattern, evaluation, and four takeaways. # Observable Channel Presentation Evidence ## Speaker Self-Identification (Explicit Claims) - **Name and role**: "I'm Joe Bergam. I'm the CEO of Hornet Dev" (transcript opening). - **Company activity**: "At Hornet we are building retrieval infrastructure for agents." - **Experience claim**: "I've been working on search and retrieval problems for a long time... more than 20 years." ## Claimed Scope - **Company scope**: Building retrieval infrastructure for agents; working with companies building web-search infrastructure: "We are also working with some of these companies to help them build infrastructure for powering that kind of use case." - **Technical positioning**: "We at hornet we are betting on BM25 as one of the primitives and we have set out on a vision to kind of have the best most efficient way to evaluate uh BM25." - **Performance claim (creator claim, unverifiable from material)**: "Hornet has a much more efficient implementation than the other engines" comparing "anonymized engines... 100 million web documents on a single node." ## Intended Audience - **Technical/developer conference audience**, evidenced by: - "All you use coding agents so you all seen GP and rip gp and said and whatever" - References to a prior speaker: "Edo here uh demonstrated what I call code mode" - Audience interaction: "how many of you heard about BM25 before?" and a clarification question from the audience ("What is the yaxis?") - Assumes familiarity with information retrieval concepts (BM25, nDCG, embeddings) while still defining BM25 basics for the mixed audience. ## Demonstrated Presentation Patterns (Observable, Not Claimed) - **Tutorial-plus-advocacy format**: Defines terms ("BM25 actually stands for best match 25"), surveys benchmarks (BrowseComp Plus), then argues a thesis ("unreasonable effectiveness of BM25"). - **Cites external research**: Waterloo / Jimmy Lynn's group paper "scaling direct corpus interaction via dynamic workspace expansion"; AOL query log comparison. - **Self-promotion of company content**: "We described it in a recent blog post as well. You can find it on hornet.dev." - **Personal, casual delivery style**: World Cup aside ("Norway is leading"), floppy disc nostalgia analogy, "if you hated it, you can tweet at me." - **Accessibility aids**: Promises to publish talk and slides; offers contact via X account. ## Audience Address and Tone - **Direct address to practitioners**: Frames takeaways explicitly: "So four claims to take away from this talk..." - **No private motivation inferred** — promotional intent for Hornet.dev is directly evidenced by the benchmark slide, blog post mention, and efficiency claims, not speculated. # Delivery Observation: Jo Kristian Bergum — "The Unreasonable Effectiveness of BM25 for Agentic Search" ## Tone & Overall Style - **Conversational, low-key, slightly informal.** Frequent filler words ("uh," "um," "right," "you know") run throughout, e.g., *"So how many of you heard about BM25 before? Is it new or is it Well, quite a few."* - **Self-deprecating age humor.** Repeated joking about his long career: *"As you can tell I'm gray hair,"* and *"I like to compare context windows with floppy discs because I'm old in the 80s, right? ... you guys are so young."* ## Recurring Phrases & Verbal Habits - **"Right?" as a constant check-in particle** — appears repeatedly: *"you come up with a score... with regards to the query. Right?"*, *"the model given the evidence up front answers the question... right?"* - **"Kind of" as a hedge/softener**, very frequent: *"kind of a scoring function," "kind of a hack," "it's kind of tiny."* - **"So" as a sentence opener** used to reset and advance: *"So BM25 actually stands for best match 25," "So exact matching is still relevant right because..."* ## Teaching Habits - **Audience polling to open.** Immediately asks the room: *"how many of you heard about BM25 before?"* and later *"Do you have pub quizzes in the US?"* - **Defining terms early.** Explicitly defines his core concept: *"my definition is that agentic search is essentially search inside agent loop."* - **Etymology/origin explanation for the technical term:** *"BM25 actually stands for best match 25... experiment number 25 turned out to be the best one."* - **Concrete analogies to familiar objects.** The floppy-disc analogy for context windows: *"one floppy disc could fit about 1.4 megabytes of data... you're still... limited to a context window that is approximately a floppy disk right."* Also frames benchmark questions as *"a pub quiz."* - **Structured enumeration.** Lists components explicitly: *"you essentially need three things... a capable model... a harness... a retrieval engine."* Closes with a numbered summary: *"Four claims to take away... That's number one. Number two..."* ## Rhetorical Devices - **Memorable coined questions as thesis lines.** *"Which BM25 do you mean?"* is introduced mid-talk (*"So I like this which BM25 do you mean?"*) and repeated verbatim in the summary — a deliberate callback. - **Paradigm framing:** casts BM25's revival as caused by a change in the user, not the tool: *"BM25 hasn't changed... the change here is really that we got a more powerful user."* The phrase **"more powerful user"** recurs at least three times. ## Communication Patterns - **Acknowledges prior speaker.** References the earlier talk: *"Edo here uh demonstrated what I call code mode."* — habit of situating his talk within the session. - **Off-topic rapport moment.** Opens with a World Cup aside: *"Norway is playing there uh against the Ivory Coast second half. Norway is leading so that's good."* - **Self-correction under audience correction.** When asked about the chart axis, he corrects himself twice: *"The y-axis is QPS. Sorry... Oh, sorry. The Y axis is latency."* - **Casual sign-off with contact invitation and humor:** *"if you hated it, you can tweet at me... happy to chat about retrieval."* - **Entertains his own speculations, flagged as such:** *"might be that when we get AGI they can just use the browser we'll see."* ## Stability vs. Isolated Moments - **Stable patterns (well-supported by repetition):** filler-heavy conversational tone, "right?" / "kind of" / "so" habits, analogy-based teaching, numbered structure. - **Isolated moments (uncertain as patterns):** the y-axis correction and the World Cup aside are one-off events; no evidence they reflect general delivery tendencies. # Position Record: "The unreasonable effectiveness of BM25 for agentic search" — Jo Kristian Bergum, Hornet.dev ## Speaker **Jo Kristian Bergum** (identified in title), CEO of Hornet.dev, self-described 20+ year search/retrieval veteran. ## Core Thesis BM25, a ~30-year-old lexical scoring function, is making a strong comeback because the "user" of search has changed — the LLM agent is a far more powerful search user than a human. ## Definitions & Framing - **Agentic search**: defined by Bergum as "search inside an agent loop" — an agent pursuing a task (coding, deep research) with an information need. - **Three requirements for agentic search**: - A capable model able to use tools and formulate queries - A harness exposing retrieval (via tool calling or "code mode," which a prior speaker "Edo" demonstrated) - A retrieval engine scaling to billions of documents - **BM25**: "Best Match 25" — named because experiment #25 performed best; a scoring function over query/document term interaction, proxying relevance. Separate from top-K retrieval acceleration algorithms. ## Key Claims & Reasoning ### 1. The change is the user, not BM25 - BM25 itself is unchanged; LLMs bring parametric general knowledge (entities, companies, dates) that makes them "very good at search." - BM25 was historically just an IR research baseline against which neural methods were compared. ### 2. BrowseComp+ benchmark findings - Deep research benchmark (~830 riddle-like questions), simple search tool over ~100K web documents, golden reference answers. - **Key observation**: If evidence is stuffed directly into the context window, accuracy is very high (even GPT-4) — "reasoning is not the bottleneck." With a retrieval harness, accuracy falls, depending on query formulation and retrieval quality. ### 3. Context window ≈ floppy disc analogy - Bergum's opinion: models degrade in quality around **350,000 tokens ≈ 1.4MB ≈ one floppy disc**. Even "AGI" models would remain context-limited, so retrieval stays relevant. ### 4. Query behavior of GPT-5 - Analysis of search trajectories (described in a Hornet blog post) shows GPT-5 writes very long queries with syntax operators (site:, phrases), unlike short human queries (AOL log comparison). "This is a new type of workload." ### 5. "Which BM25 do you mean?" - The BM25 baseline in the BrowseComp+ paper used hyperparameters inadequate for long documents; recent research shows this dramatically affects results, making fancier techniques look better than they should. ### 6. Why BM25 suits agentic search - Exact matching suits the model's knowledge of names, entities, zip codes, SKUs — hard for embedding models with fixed vocabularies - Cheap vs. embedding inference (some embedding models ~8B parameters) - Good ecosystem tooling - **Explainable to the model**: literal term matching lets the model understand results and reformulate queries ### 7. Dynamic workspace expansion (Waterloo, Jimmy Lin's group) - Paper: "Scaling direct corpus interaction via dynamic workspace expansion" — Bergum says he loves it. - Retrieval results placed in a file-system workspace; progressive disclosure (title + snippet, model reads more as needed); combines sandbox, VFS, bash, retrieval infrastructure. - Bergum frames this as a **"hack"** aligned with frontier-lab optimization for coding/tool use — models improve on this trajectory automatically. Tentative: "might be that when we get AGI they can just use the browser, we'll see." ### 8. Evaluation - Classical single-query IR evaluation (nDCG on one ranked list) is "now kind of dead"; evaluate end-to-end task success (e.g., does it answer correctly). - Single-shot query evaluation less relevant since agents issue many more queries than humans. ### 9. Hornet performance claim - Anonymized engine comparison: 100 million web documents, single node, same hardware — Hornet is more efficient with more throughput, implying cost savings. ## Contradiction / Correction Preserved - When asked about the chart's y-axis, Bergum first says **QPS**, then corrects himself: **"Oh, sorry. The Y axis is latency."** The corrected claim stands as stated, with the initial misstatement preserved here. ## Stated Qualifications & Tentativeness - 350K-token degradation figure is explicitly "in my opinion" - AGI/browser speculation marked with "we'll see" - "Currently this is a very powerful way" — implies possibly temporary advantage ## Closing Takeaways (per Bergum) 1. **New user**: faster typing/reading, query reformulation, general knowledge — makes simple tools (grep + BM25) more powerful 2. **Which BM25?** Implementations, performance, and parameters differ — matters 3. **Why effective**: explainable to the model; combines with grep since "grep is almost about literal matches as well" — the combination is a strong agentic retrieval paradigm # Speaker Attribution & Provenance Record ## Video - **Title:** The unreasonable effectiveness of BM25 for agentic search — Jo Kristian Bergum, Hornet.dev - **Video ID:** `01a15ca7-602f-7d1b-b4a9-4291a26a9f58` ## Speakers & Roles ### Primary Speaker - **Name:** Jo Kristian Bergum (introduces himself as "Joe Bergam" in transcript) - **Role:** **Creator / Presenter** — CEO of Hornet.dev - **Self-described credentials:** 20+ years working in search and retrieval ### Secondary Voice(s) - **Audience member(s):** At least one unnamed audience member asks a clarifying question about a chart ("What is the y-axis?"); speaker first answers "QPS," then corrects to "latency." - **Referenced third party:** "Edo" — cited (not speaking) as having demonstrated "code mode" for exposing retrieval infrastructure earlier in the event. ## Referenced Sources & Claims - **BrowseComp Plus** — deep research benchmark, ~830 riddle-like questions, ~100,000-document corpus of web documents; cited for end-to-end accuracy findings and as an example of a poorly parameterized BM25 baseline in the original paper. - **Waterloo research — Jimmy Lin's group** — paper "Scaling Direct Corpus Interaction via Dynamic Workspace Expansion"; cited on agentic search and workspace/VFS approaches. - **AOL query log** — cited as historical comparison for query length patterns vs. GPT-5 long-form queries. - **Hornet blog post** — analysis of GPT-5 query trajectories on BrowseComp Plus, said to be available at hornet.dev. - **GPT-4, GPT-5** — referenced models, not sources of claims. ## Promotional Segments (Flagged) - **Hornet.dev positioning:** Speaker is CEO of Hornet.dev; describes company as "building retrieval infrastructure for agents" and working with companies building web search infrastructure. - **Benchmark comparison slide:** Anonymized competitor engines vs. Hornet throughput/latency on 100 million web documents, single node — a performance claim for the speaker's own product, including a mid-presentation correction about the y-axis metric. ## Availability Limits - **No curated metadata** was supplied; provenance facts derive solely from the transcript. - **No Q&A session:** Speaker states there is "not a room for questions"; offers contact via X (Twitter) account and in-person conversation. - **Slides/references:** Speaker says the talk and slides "will be published" — not available in this transcript. - **Comment availability:** Not stated in the source material. So great being here. Uh I'm Joe Bergam. I'm the CEO of Hornet Dev and I'm here today to talk about the unreasonable effectiveness of BM25 for Agentic Search. So how many of you heard about BM25 before? Is it new or is it Well, quite a few. So that's great. Uh I'm also watching the World Cup. Norway is playing there uh against the Ivory Coast second half. Norway is leading so that's good. So yeah and at Hornet we are building retrieval infrastructure for agents and uh I've been working on search and retrieval problems for a long time. As you can tell I'm gray hair. Uh been working in this space for more than 20 years. Um and in the talk today I'll talk about why this kind of 30-year-old lexical scoring function is uh making a strong comeback. First I will talk a little bit about what I mean by agentic search or aentic dribble and kind of define that for for you all of you. Um my definition is that agentic search is essentially search inside agent loop. So you have an agent that is trying to accomplish a task, write a coding or do some deep research or whatever task and inside that you have some kind of information need for the agent, right? In order to get that task done successfully and you essentially need three things uh to kind of build a good agentic search system and that is you need a capable model, a model that is able to use tools and is able to formulate queries. You also need a harness around the model and how you kind of expose the retrieval and the search functions to the model. There's different ways to do that. It could be through tool calling or it could be through code mode. Um, Edo here uh demonstrated what I call code mode for exposing retrieval infrastructure. So that's the harness part and you also need uh a retrieval engine to be able to do searches efficiently potentially over billion scaled uh document sets and to define BM25. So BM25 actually stands for best match 25. So there were some researchers doing plenty of different experiments and experiment number 25 turned out to be the best one. So that's that's the background of the name. It's essentially a scoring function. So you can imagine you have a query and you have a document and you calculate some kind of score by interaction between the query terms and the document terms and you come up with a score and you hope that this score kind of is a good proxy for the relevancy of the document with regards to the query. Right? So one way to calculate BM35 would be to take all the documents and score each of them you know and then figure out what are the top K documents and then there's like 30 or 40 years of interest in how you accelerate that type of retrieval top K lots of different algorithms we also invest a lot in that I'll show something in this direction but BM25 is the scoring function and there is a way to kind of accelerate top K retrieval Uh BM25 hasn't changed. It's the same scoring function, but the change here is really that we got a more powerful user. Uh Edo talked about general knowledge. The LLMs today have a lot of general knowledge. They know about entities, they know about companies, they know about dates, they know a lot, right? And by using that kind of implicit knowledge that is built into the parametric model, uh they essentially become very good at search. Um and that's the really change uh here now that we're is kind of making BM25 more relevant and BM25 used to be a kind of a baseline function. Any information retrieval research would include a BM25 baseline and then you would put something fancy advanced neural fancy stuff and then you would compare it with BM25. I also think it's interesting in how we evaluate search before because you will simply look at 10 blue links and you like have scan it and you compute some metrics. A lot of that is now going away because the agent is not really it's kind of powerful in the way that it can type out a lot more queries than a human can do. So it's less relevant to think about evaluating these systems just by a single shot query. And this is one of my kind of favorite benchmarks out there. I like to talk about benchmarks. So browsecom plus is a deep research benchmark \[snorts\] uh published in a paper last year and it has almost or exactly 830 questions. These are riddle like think about it like a pub quiz. Do you have pub quizzes in the US? Yeah. Okay great. So like kind of a riddle type of questions quite long and the agent the harness of this kind of or the protocol of this benchmark is that you have a model and it gets a very simple tool called search and it accepts a query string and you return some snippets back to the model and the corpus is about 105,000 or 100,000 documents. So that's kind of tiny and these are web documents and the end to end accuracy. Um all of these questions have a golden reference answer and you can kind of check if the model and the entire loop produces that exact answer. Um but why kind of why do we need retrieval? I like to compare context windows with floppy discs because I'm old in the 80s, right? We install this kind of games on our computers using floppy discs. So one kind of you guys are so young so you don't probably have this kind of nostalgy but one floppy disc could fit about 1.4 megabytes of data and the current models before they start degrading in quality that's I in my opinion around 350,000 tokens. So that's one floppy disc of data right. Um, so you need retrieval in order to to fetch the information that you actually need to put into the context window and browse plus really demonstrate how uh retrieval quality affects the end toend accuracy of the task. Right? The endto-end accuracy here is essentially is the model equipped with this search tool able to answer the question, right? the riddle like question and if you artificially just stuff the evidence documents that is needed to answer this question into the context window of the model the accuracy is really high right so reasoning is not the bottleneck the model given the edit evidence up front answers the question with a very high accuracy rate even GPT4 um but if you expose the model with a harness with a retrieval tool. That accuracy falls because it now depends on the harness. It depends on the model's ability to formulate queries and the retrieval quality of the retriever. Um so for me this is also quite important because even if we get perfect models like models that are kind of agi you don't have to append make no mistakes you still will be limited to a context window that is approximately a floppy disk right so you have to decide what goes into that context window and I think retrieval is still very relevant as the previous slide showed and in this browsecom plus data set one of these riddlelike questions become a search trajectory because the model execute a query, gets some response back, reads it, reformulates the query and continues until it has kind of filled up the context window or found the answer, whatever comes first. And we spend some time to investigate these trajectories uh to see how GPT5 is formulating queries and we found a lot of interesting aspects with that. Uh we described it in a recent blog post as well. You can find it on hornet.dev. And we like to compare it with a AOL query log. So AOL AOL was like a service back in the day had some search interface and they accidentally published um a very large sample of what people were searching for on the web and they were quite short and I have seen more recent query logs as well and the user human pattern are still searching with just a few terms. GPD5 on the other hand it's a much more powerful user. It has the general knowledge and it can like bam write out very long queries. It can use a lot of syntax operators that are kind of useful from it has learned from web search. So site operator phrases etc. And this is a new type of workload. And on BM25, BM25 has essentially two hyperparameters that controls various aspects of the scoring function. And I talked about having a baseline and BM25 is usually a baseline and browser comp plus as well has a baseline with BM25, but it turns out that that baseline is terrible. So when you look at fancier techniques, embedding models, what have you, um it's stands out as a much better retrieval paradigm than BM25 if you look at the original paper. But more recent research shows that the parameters that were used in the browse comp plus research paper was not really adequate to handle these kind of long documents. So I like this which BM25 do you mean? uh because it has a quite dramatic impact on that specific config benchmarks on the overall accuracy and why is BM25 now more powerful with the new user. So I mentioned the general knowledge of the user and that he can type faster and more be more specific as a more powerful user and exact matching is still relevant right because the model knows names entities uh zip codes uh skus what have you that is not so easy to represent with an embedding model which kind of encodes all the tokens into a fixed vocabulary. Uh it's also relatively cheap especially if you take into consideration the cost of doing embedding inference right some of these embedding models have like 8 billion parameters and you encode text and you have to stand up infrastructure for this and what have you. So it's quite simple and also the tooling uh in the overall ecosystem is quite good. Um, so it's it's readily available and it's also very easy for the model to inspect the results and understand why a certain query formulation returned the result it did, right? Because you're matching literal terms and phrases and things like that which can help it kind of reformulate the queries. So these are the three key things and now into more hot topics on like what is all you need. Um and this is a very recent research that came out from Waterloo from Jimmy Lynn's group up there. They are doing a great job at the information retrieval research and also on agentic search. They have a recent paper that I love. Uh it's called scaling direct corpus interaction via dynamic workspace expansion. So I'll spend some time on expanding this. So imagine you want to stand up web search infrastructure for agents. A lot of companies are doing that at the moment. We are also working with some of these companies to help them build infrastructure for powering that kind of use case. And there you have potentially billions of documents, right? So that doesn't fit into the context window. So you obviously need retrieval and BM25 is a good baseline. So you can retrieve information over that. And the result of this is um you can imagine this as a search engine result page SER for agents because you can place these documents that are retrieved from the retriever into a workspace. And if you organize this workspace as a file system, you can uh play into the same things that you have around skills. You can have progressive disclosure because you can have the document like the title of the document and a small snippet of the document and expose that to the model and the model can then also decide oh I need to read more of the document and when it's doing that it can use all the primitive tools that it's really good at using you all use coding agents so you all seen GP and rip gp and said and whatever it's doing to kind of manage context so here you get the benefits of both worlds and also So you get to combine sandbox infrastructure, retrieval infrastructure. So and VFS and just bash and whatever. So it's also quite exciting, right? Because it combines all of these new type of paradigms that is happening at the moment. So I'm really excited about this type of direction. And it's also kind of a hack, right, to optimize for what the models are good at at the moment, right? Because all the frontier LLM companies are optimizing their model for cool for coding, for bash, for tool use. So if you put your kind of end to end task on the trajectory of that you kind of whatever there's a new model you know there's going to be better at this as well right so might be that when we get AGI they can just use the browser we'll see but currently this is a very powerful way to to build uh retrieval infrastructure and a whole agentic search experience and when it comes to evaluation right I talked about this earlier as well. In traditional information retrieval, we were used to having just one query, one rank list, compute NDG and compare it. That is no longer very relevant when the new user is agent because that agent can reformulate queries and do more queries and do expansion and all kinds of different stuff, right? So, a lot of the classical information retrieval evaluation is now kind of dead. instead look at like see if the model can perform the task it's it's set to and for example for question answering does it get the answer right and we at hornet we are betting on BM25 as one of the primitives and we have set out on a vision to kind of have the best most efficient way to evaluate uh BM25 because I think it's such a strong fundamental primitive And this illustration compares some anonymized engines comparing with Hornet on the same type of same type of hardware with um web documents, 100 million web documents on a single node. And as you can see, Hornet has a much more efficient implementation than the other engines and can do more throughput for the same type of bucks, which for a lot of companies that are building infrastructure at the moment for web search etc. means a lot in savings. \>\> What is the yaxis? \>\> The y-axis is QPS. Sorry. \>\> The Y axis is Oh, sorry. The Y axis is latency. I'm sorry. So, four claims to take away from this talk. There's a new user. It's more powerful. It's able to type faster. It's able to read faster. It's able to reformulate queries and it has a lot of general knowledge which makes simple tools like GP and BM25 more powerful. That's number one. Number two, which BM25 do you mean? Uh there are differences in implementation, in performance, in parameters. So think about that. And also why is it effective for agentic search? It's simply because it's explainable for the model. So the model can see and also use it in combination with GP, right? Because you have literal matches and GP is almost about literal matches as well. And the combination of these two is a very strong agentic search or agentic retrieval paradigm. There's a lot of references. I think I will publish a talk uh or the talk will be published and also the the slides and if you hated it, you can tweet at me. Um there is not I was told that there's not a room for for questions but happy to chat about retrieval. You will find me around the conference. Probably best way to reach me is through my ex account and that's it.