JP Sherman - Reciprocal Rank Fusion: How a Query’s Intents Collaborate to Build an Answer

1. TECH SEO CONNECT · DURHAM · NOVEMBER 5–6, 2026 Reciprocal Rank Fusion How a query becomes an answer JP Sherman Principal Product Manager, Search, Red Hat
2. BEFORE THE QUERY JP Sherman Principal Product Manager, Search, Red Hat Staff Sergeant, Civil Air Patrol (AUX USAF) Aerial Search & Rescue 15 years on search engineering teams 15 years on technical SEO teams U.S. Army PSYOP (Airborne) Tech SEO Connect · Durham · November 5–6, 2026 02 / 22
3. THE WORDS Eight terms. We will meet each one as the query moves. Query Intent The words a person typed. The job inside those words. A hang, a registration, a trial. Index Fan-out The pages kept together for one kind of question. One query, written as several searches. List, or ballot Place One search, returned in order. These two words mean the same thing. Where a passage sits on a list. The math keeps this number. Window Passage How far down a list the count looks. The piece that can take a place. One page can hold several. Tech SEO Connect · Durham · November 5–6, 2026 03 / 22
4. THE WHOLE PATH A query goes in. An answer comes out. 1 2 3 4 5 6 Query (Red Hat Homebrew) Gateway Fan-out RRF Clef Answer Several searches, several lists The places are added The odds it can be quoted A person can use it The words they typed The intent picks an index Eight lists go into the fusion. One sentence comes out. Reciprocal rank fusion adds the places. Clef returns the odds. A person judges the answer. Tech SEO Connect · Durham · November 5–6, 2026 04 / 22
5. WHERE RANKING STARTS We start with the query. 1 HUMAN 2 ROBOT 3 HUMAN A person has a need and types it. The query goes through six steps and comes out as an answer. A person reads the answer and judges it. This happens outside the ranking system. The query is the human part we are given. First a conductor sends it to the right index. Then the six steps. Our pages are what it compares the query against. Did it solve the need they started with? That judgment sits outside the ranker too. Human, then robot, then human. We follow the query from the person who typed it, through the six steps, to the person who evaluates the answer. Tech SEO Connect · Durham · November 5–6, 2026 05 / 22
6. INTELLIGENT RETRIEVAL GATEWAY A conductor for a house with many indexes. The person has typed a query. Troubleshooting passages live in one index. Events live in another. Trials live in a third. The gateway reads the intent and sends the query to the index that holds it. What it reads What it matches What it passes on The intent of the query. A hang. A registration. A trial. The index built for that intent. The other indexes stay on their own shelves. That query, aimed at that index, into reciprocal rank fusion. The gateway chooses the index. Fusion is still how the lists on that shelf are added. The conductor does not add the places. Tech SEO Connect · Durham · November 5–6, 2026 06 / 22
7. THE ROUTE One query, sent to the index that holds its intent. “the install hangs” the query INTELLIGENT RETRIEVAL GATEWAY Reads the intent. This one is troubleshooting. Troubleshooting Events Trials Crashes, hangs, support. This index. Registrations and the summit. Evaluations and trying the product. On to reciprocal rank fusion. The lists that come back from the troubleshooting index are the lists the math will add. Tech SEO Connect · Durham · November 5–6, 2026 07 / 22
8. THE ROBOT Six steps turn the query into an answer. 1 2 3 Recognize Understand Fan out Decide what kind of need this query is, before any page is retrieved. Fill in the entity and the constraint the person did not type. Turn this one query into several searches. 4 5 6 Retrieve Fuse Decide Compare those searches with the passages on your site. Merge those lists by place, not by raw score. Keep a passage only if it can answer the query. Tech SEO Connect · Durham · November 5–6, 2026 08 / 22
9. WHAT FAN-OUT LOOKS LIKE One query splits into many searches. What is open source? a definition Which license applies Open source vs paid the rules a comparison open source software the query Is it safe to use? Best tools to start a risk a shortlist How to contribute a task Tech SEO Connect · Durham · November 5–6, 2026 09 / 22
10. A BALLOT One search comes back as one ranked list. THE SEARCH open source software 1 The Words by Themselves The results, like those pulled from SOLR or other algorithmic based search engines start with the words in orders. The Exact Words Meaning & Intent is Then Applied 2 The Meaning of the Phrase 3 Re-Rankers and Contextualizers Vectorization allows meaning to be inferred from similar words, phrases, ideas and contexts giving a more wholistic view of the intent. Re-Ranking Contextually With the rank scores from just the exact and vectorized words, a third HYBRID way is a way to re-trank them considering both of these things at the same time. This is the Fusion in Recriprocal Rank Fusion Tech SEO Connect · Durham · November 5–6, 2026 10 / 22
11. THE FOUNDATION Four spokes become eight ballots, and eighty places. 4 × 2 = 8 × 10 = 80 spokes searches each ballots places kept places What is it Vs paid Is it safe How to start Exact words Exact words Exact words Exact words The idea The idea The idea The idea 80 places on the lists. Not 80 different passages. The same chunk can take a place on more than one ballot, and one URL can hold more than one chunk. Place 11 on any list adds nothing. A deeper cutoff, 50 instead of 10, would be 400 places. The ballots would still be 8. Tech SEO Connect · Durham · November 5–6, 2026 11 / 22
12. THE BALLOTS ARE IN Now the math decides which pages get picked. One page. Three ballots. The bars are what happens if you try to add the scores. 14.2 Exact words The Vector (Idea) 0.83 The Re-Rank 2.7 Same page, kept as a place on each ballot. 4th 1st 9th Exact words The idea The Re Rank The loud score was not first. Best match. Score was 0.83. Refines the output 14.2 swallows the other scores, and it was only 4th. Throw the scores away. Add the places. The bars above cannot be added. The three places can. Tech SEO Connect · Durham · November 5–6, 2026 12 / 22
13. THE RULE Add 60 to the place. Then divide 1 by it. FIRST PLACE TENTH PLACE 60 + 1 = 61 1 / 61 = 0.0164 60 + 10 = 70 1 / 70 = 0.0143 The whole vote from one ballot. Nine places later. Almost the same bar. 0.0164 − 0.0143 = 0.002 Climbing from tenth to first barely moves the vote. Showing up on a second list changes the selection more. Tech SEO Connect · Durham · November 5–6, 2026 13 / 22
14. WHY THE 60 IS THERE This prevents popular content from being popular because it is popular. IF THE 60 WERE 0 WITH THE 60 1st place = 1 / 1 = 1.00 1st place = 1 / 61 = 0.0164 2nd place = 1 / 2 = 0.50 2nd place = 1 / 62 = 0.0161 First place is worth two second places. The two bars are almost the same length. 60 was chosen once and left there. Elasticsearch uses 60 unless someone changes it. You can explain the selection without tuning the number. Tech SEO Connect · Durham · November 5–6, 2026 14 / 22
15. ADD THE BALLOTS The selection is the total, not the best place. WON ONLY ONE LIST GOOD ON THREE LISTS Exact words 1st 0.0164 Exact words 12th 0.0139 The idea absent 0 The idea 8th 0.0147 A third list absent 0 A third list 20th 0.0125 TOTAL 0.0164 one vote TOTAL 0.0411 0.0139 + 0.0147 + 0.0125 2.5× Nobody on the right was first. Three modest places still beat one first-place finish, because the missing lists added zero. Tech SEO Connect · Durham · November 5–6, 2026 15 / 22
16. THE CUTOFF Only the top of a list gets a vote. Each ballot is cut off. A common default keeps the first 10 places and drops the rest. Past that line, the vote is zero. 1 2 3 4 5 6 vote vote vote vote vote vote 7 8 9 10 11 12 vote vote vote vote zero zero Place 11 adds nothing. It is not a small vote. It is the same as a page that never appeared on the list. Elasticsearch uses 10 unless someone changes it. Some systems keep 50 or 100, and only then can a lower place count. Classic technical SEO is this gate. Crawl, index, render, canonicals, and a passage the list can score are what get a page into this top. Nothing later repairs a page the list never kept. Tech SEO Connect · Durham · November 5–6, 2026 16 / 22
17. AFTER THE COUNT The lists built an order. They did not choose a quote. THE LISTS THE EDITOR Places, added. Odds, on one passage. Reciprocal rank fusion builds this order. A passage rises by showing up on more ballots. That sum stays. You send Jev or Clef the query, this passage, and questions you wrote. They return a probability for each question. Your code keeps or drops the passage. The formula stays 1 divided by 60 plus the place. The checklist is a later call, on passages the lists already kept. Your code chooses the quote. Tech SEO Connect · Durham · November 5–6, 2026 17 / 22
18. JEV One sentence. Three questions about it. Jev tells you the odds. THE SENTENCE Yes or no Does this say what open source software is? “Open source software can be changed. The license sets the limit.” You get the odds of yes. You send this sentence, and the questions on the right. Pick one Jev writes nothing. Jev outputs a likelihood. Definition, comparison, or downloading bits? You get one label, and the odds for each. How far Can a person quote this sentence alone? You get a mark on a scale you wrote. Tech SEO Connect · Durham · November 5–6, 2026 18 / 22
19. CLEF The same three questions. Clef understands a wider format Yes or no Pick one How far Sees Holds Yours Images and video 64K · 64 questions Apache 2.0 Flash Clef Jev 35.6 79.4 76.5 Flash is the open-source, multimodal AI decision models NOT THAT Flash that makes TECH SEOs twitch Tech SEO Connect · Durham · November 5–6, 2026 19 / 22
20. AFTER CLEF The highest total can still be the wrong quote. The places have already been added. Clef reads the passage and tells you whether it can be the answer. WON THE LISTS QUIET ON THE LISTS “Open source is popular and widely used.” “Open source software can be changed. The license sets the limit.” First on several ballots. The sentence never says what open source software is. Only decent on the ballots. The sentence says the thing and the limit. Not quoted. Quoted. Clef does not change the totals. It chooses which sentence you show. Tech SEO Connect · Durham · November 5–6, 2026 20 / 22
21. TECHNICAL SEO OK Nerd. What does this have to do with TECH SEO? 1 2 3 Obsess (in a normal amount) About Crawl Budgets, Optimization & Indexation IA, Schema, Nav and On Page Content Work Closer Together Information architecture extends inside the URL. One section for each meaning. Put the words of that search and the fact in the same sentences, so both listeners can vote for it. Build coherent sections that keep answers, entities, conditions, and evidence together and/ or linked. A retrieval system may use a passage even when the user ultimately lands on the page. Crawl it, index it, render it, and give it a canonical. The passage has to be one a list can score. Place 11 adds nothing. You build the passages. The ranker can only use the ones it can find. Tech SEO Connect · Durham · November 5–6, 2026 21 / 22
22. AFTER CLEF Rethinking the URL in the Traditional Tech SEO Model The URL remains the publishing address. It is no longer a sufficient model of the information a search system can retrieve and use. We need to design the page and the meaningful passages inside it. Content processing Retrieval Rank fusion Whether the system represents pages, passages, entities, or other units Which candidates enter each ranked list How those lists are combined Tech SEO Connect · Durham · November 5–6, 2026 20 / 22
23. AFTER CLEF Clarity and Specificity in Content Look for opportunities to add clarity and context - you can assume that the user may not be reading it in your environment consider a fictional product guide. “It supports this configuration, but only on the newer version.” Lacks clarity and context if extracted from the source. “Acme Gateway 4.2 supports SAML single sign-on on Enterprise plans. Version 4.1 supports password authentication only.” Ready for extraction. Quoted. Humans are still particularly stronger for more nuanced writing. Tech SEO Connect · Durham · November 5–6, 2026 20 / 22
24. AND THAT IS HOW I BECOME AN ANSWER You structure the information. Retrieval determines what enters the lists. Fusion determines how those candidates combine. Your IA is that library. JP Sherman Tech SEO Connect · November 2026