--- title: "How to Weight Vector vs Keyword Search" url: https://www.insulin.dev/blog/tune-hybrid-search-vector-keyword-weight/ canonical: https://www.insulin.dev/blog/tune-hybrid-search-vector-keyword-weight/ type: Blog description: "Hybrid search has a dial. How to weight vector vs keyword matching in a knowledge base, when to pick a pure mode, and how to read the relevance score." --- # How to Weight Vector vs Keyword Search > Canonical HTML version: https://www.insulin.dev/blog/tune-hybrid-search-vector-keyword-weight/ 1. [Home](/) 2. / 3. [Blog](/blog/) 4. / 5. How to Weight Vector vs Keyword Search # How to Weight Vector vs Keyword Search Hybrid search has a dial. How to weight vector vs keyword matching in a knowledge base, when to pick a pure mode, and how to read the relevance score. ![Zee Chen](/authors/zee-chen.jpg) Zee Chen Aug 18, 2026 ![How to Weight Vector vs Keyword Search](/images/blog/tune-hybrid-search-vector-keyword-weight/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Ftune-hybrid-search-vector-keyword-weight%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Search%2C%20Knowledge%20Bases. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Ftune-hybrid-search-vector-keyword-weight%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Search%2C%20Knowledge%20Bases. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Ftune-hybrid-search-vector-keyword-weight%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Search%2C%20Knowledge%20Bases. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Ftune-hybrid-search-vector-keyword-weight%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Search%2C%20Knowledge%20Bases. "Summarize with Perplexity") Table of Contents - [What the vector weight actually controls](#what-the-vector-weight-actually-controls) - [When to shift the dial toward vector](#when-to-shift-the-dial-toward-vector) - [When to shift the dial toward keyword](#when-to-shift-the-dial-toward-keyword) - [The two pure modes, and when to use them](#the-two-pure-modes-and-when-to-use-them) - [Read the relevance score to prove it worked](#read-the-relevance-score-to-prove-it-worked) - [A weight is per base, and an agent reads up to three](#a-weight-is-per-base-and-an-agent-reads-up-to-three) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _Hybrid search combines vector similarity with keyword matching, and you choose how much weight the vector side carries. Here is how to set that dial, when to drop to a pure mode, and how the test-search relevance score tells you whether you got it right._ * * * Most guides on knowledge-base search stop at “use hybrid, it’s the best of both.” That is true, and it is also where the real question begins. Hybrid is not one setting — it is a dial, and where you put the dial changes which passages come back first. A knowledge base of contract clause numbers wants a different setting than one of loosely-worded runbooks, and the default may serve neither perfectly. This is the tuning guide. Not _what_ hybrid search is — [that comparison lives here](/blog/hybrid-search-vs-vector-search/) — but how to weight it, when to abandon hybrid entirely, and how to prove your choice was right before an agent ships answers off it. * * * ## **What the vector weight actually controls** The vector weight is a single number that decides how much a knowledge base favours meaning over exact terms when it ranks results. In Insulin, [a knowledge base’s search](/knowledge-bases/) runs in one of three modes — Hybrid, Vector only, or Keyword only — and in Hybrid mode the vector weight sits on a 0-to-1 scale: how much to favour vector matches over keyword matches. The product page states it plainly: “Hybrid search combines vector similarity with keyword matching, and you choose how much weight the vector side carries.” That choice is the whole game. A low weight ranks the passage that contains your exact string; a high weight ranks the passage that means what you asked, even if it shares no words with the question. The two matching methods do not compete so much as cover for each other. Keyword matching finds `FY26-DISC-03` when someone types `FY26-DISC-03`. Vector matching finds the discount policy when someone asks “the thing finance has to sign off on.” The weight decides which instinct wins when both find something. * * * ## **When to shift the dial toward vector** Shift toward vector — a higher weight — when your readers and your documents use different words for the same thing. This is the common case for policy and procedure content, where the author wrote in precise internal language and the asker describes the idea loosely. Turn the dial up when: - **Questions are phrased in plain language.** A new hire asks “how much can I discount before someone has to approve it,” and the document says “thresholds exceeding 25% of list price require VP sign-off.” No shared keywords; only shared meaning. - **The corpus is narrative.** Runbooks, playbooks, onboarding guides, support procedures — prose that explains rather than tabulates. Meaning-matching finds the right paragraph even when the phrasing drifts. - **You care more about recall than precision.** A slightly-off but relevant passage is better than nothing when the alternative is a keyword miss that returns an empty answer. The trade is real: push the weight high enough and the search will return something _about_ the topic instead of the exact clause you named. That is fine for “explain our refund stance,” and wrong for “what does clause 7.2 say.” * * * ## **When to shift the dial toward keyword** Shift toward keyword — a lower weight — when your documents are dense with identifiers that carry no meaning to embed. Product codes, ticket numbers, SKUs, clause numbers, error strings, and internal acronyms are strings, not concepts, and vector search handles them worst. Turn the dial down when: - **The corpus is reference material.** Approval matrices, pricing tables, SKU catalogues, contract templates — content where the exact token _is_ the answer. - **Queries name a specific artefact.** “Pull up policy `FY26-DISC-03`” or “what’s the fix for error `E4021`” should return that exact thing, not its semantic neighbours. - **A near-miss is dangerous.** In a legal or financial base, retrieving the clause _next to_ the one you asked for is not a helpful approximation — it is a wrong answer wearing a citation. If a base is almost entirely identifiers and exact terms, you may not want any vector influence at all — which is what the pure modes are for. * * * ## **The two pure modes, and when to use them** Vector-only and keyword-only are the ends of the dial made explicit — a knowledge base set to search on meaning alone, or on exact terms alone. They exist for corpora that lean so hard one way that the other method only adds noise. **Keyword-only** suits a base that is nothing but structured identifiers: a parts catalogue, a table of error codes, a policy index keyed by number. There is no meaning to embed, so vector matching would only surface plausible-looking wrong rows. **Vector-only** suits a base of pure narrative where exact wording almost never matches: a library of long-form writing, meeting notes, transcripts. Keyword matching there mostly finds stopwords, so the vector side is doing all the useful work anyway. For nearly everything in between — the mixed policy-and-prose documents most teams actually load — hybrid with a tuned weight beats either extreme. Reach for a pure mode only when you can describe the corpus in one word: _identifiers_, or _prose_. * * * ## **Read the relevance score to prove it worked** The relevance score is how you check your tuning without shipping it to an agent first. A test search in a knowledge base returns the matching content, its source document, and a relevance score per result — so you can see not just _what_ came back but how strongly the search believed each result answered the query. Use it as a diagnostic, not a vanity metric: 1. **Ask a question you already know the answer to.** Pick one where you know exactly which document and passage should win. 2. **Read the ranking, not just the top hit.** The right passage should sit at the top with a clear score lead. If it is third, or the scores are bunched together, retrieval is muddy and your weight is likely off. 3. **Adjust and re-test.** If the correct passage is exact-term-based and ranked low, drop the vector weight. If a loosely-phrased query missed the obvious document, raise it. Re-run the same test search and watch the score move. This closes the loop that grounding depends on. [Grounding an agent in your documents](/blog/how-to-ground-an-ai-agent-in-your-documents/) only works if the search underneath surfaces the right passage — and the relevance score is the one place you can watch that happen before any answer is generated. * * * ## **A weight is per base, and an agent reads up to three** Because the weight lives on the knowledge base, you can tune each base to its own content and let an [Insulin agent](/agents/) draw on several at once. An agent can search up to three knowledge bases in a single turn, citing the source document for each result — so a deal-desk question might pull from a keyword-tuned pricing table and a vector-tuned policy base in the same answer. That is the practical payoff of tuning per base rather than globally. Your identifier-heavy reference base runs low on vector weight, your narrative policy base runs high, and the agent that reads both gets the right retrieval from each without a compromise setting that half-serves either. * * * ## **Frequently asked questions** **How do I weight vector vs keyword search in a knowledge base?** Set the vector weight on a 0-to-1 scale in Hybrid mode. A higher weight favours meaning-based matching, a lower weight favours exact-term matching. The weight is set per knowledge base, so tune each base to its own content. **When should I raise the vector weight?** Raise it when readers and documents use different words for the same thing — narrative content like runbooks and policies where questions are phrased loosely. A higher weight finds the right passage even when it shares no keywords with the question. **When should I use vector-only or keyword-only mode?** Use keyword-only for a base of pure identifiers — error codes, SKUs, clause numbers — where there is no meaning to embed. Use vector-only for pure narrative where exact wording rarely matches. For mixed documents, tuned hybrid beats either. **What does the relevance score tell me?** A test search returns the matching content, its source document, and a relevance score per result. Ask a question you know the answer to and check whether the correct passage ranks first with a clear score lead. If not, adjust the vector weight and re-test. **How many knowledge bases can an agent search at once?** An agent can search up to three knowledge bases in a single turn, citing the source document for each result. Because the weight is set per base, one answer can combine a keyword-tuned base and a vector-tuned base. * * * ## **Takeaways** - The vector weight is a 0-to-1 dial on a hybrid base: high favours meaning, low favours exact terms. - Shift toward vector for loosely-phrased questions over narrative documents; shift toward keyword for identifiers, tables, and named artefacts. - Reach for vector-only or keyword-only only when the corpus is describable in one word — _prose_ or _identifiers_. - The weight is set per base, and an agent can search up to three bases in a turn, so tune each to its own content instead of compromising globally. - Prove your setting with a test search: ask a question you can grade, read the relevance score, and adjust until the right passage wins clearly. The dial is only useful if you watch it move. Load a base, run a test search on a question you already know, and tune the weight until the relevance score puts the right passage first. [Explore Insulin knowledge bases](/knowledge-bases/), or [get a demo](/schedule-demo/) to tune one on your own documents. ## Sources Primary sources for the platform rules cited above. Last verified August 18, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Suger Insulin docs: Knowledge Base](https://doc.suger.io/insulin/knowledge-base/) — Search mode is Hybrid, Vector only, or Keyword only; vector weight favours vector over keyword on a 0-1 scale; an agent can search up to three knowledge bases per turn; a test search returns the content, its source document, and a relevance score. ### Stay Updated Get the latest Cloud GTM insights, product updates, and marketplace strategies delivered to your inbox.