Hybrid search combines vector similarity with keyword matching so a knowledge base finds the right passage whether the question used your internal jargon or described the idea in other words. In Insulin you choose how much weight the vector side carries.
Retrieval is the part of a grounded AI system that decides everything downstream. If the wrong passage comes back, the model will answer fluently from the wrong passage — and the citation will make that answer look verified.
So it is worth understanding the two ways a search can find a passage, and why business documents in particular need both.
The two retrieval methods
Keyword search matches the characters you typed. It is exact, predictable, and completely literal. Search for SOC 2 and you get documents containing SOC 2. Search for security certification and you do not, even though the document you want is about exactly that.
Vector search matches meaning. Documents and queries are converted into numerical representations, and the search returns passages whose representation sits close to the query’s. Ask “what security certifications do we hold” and it can surface a document that says SOC 2 Type II without those words ever appearing in your question.
Each fails in a way the other does not. Vector search is weak precisely where language is arbitrary — product codes, ticket numbers, internal acronyms, a policy called FY26-DISC-03. Those strings carry no semantic meaning to embed, and a vector search will happily return something about discounting instead of the named policy. Keyword search is weak wherever the asker and the author chose different words, which in practice is most of the time.
Why business documents need both
Corporate documents are unusually bad for pure vector search, because they are dense with exactly the tokens that embeddings handle worst.
An approval matrix is full of thresholds and role names. A runbook is full of system names, field names, and error strings. A contract template is full of clause numbers. Meanwhile the people asking questions describe things loosely — “the thing where finance has to sign off” — because they do not know the internal name. That is the case for vector search.
You need retrieval that can do both in one query: find FY26-DISC-03 when somebody names it, and find it when somebody describes it. That is what hybrid does — it runs both and merges the results, so a passage that either method finds strongly is a passage that surfaces.
Tuning the weighting
Insulin lets a knowledge base run hybrid (the recommended default), vector-only, or keyword-only, and on hybrid a vector weight from 0 to 1 sets how much the vector side is favoured over keyword matching. The default is a reasonable starting point; the useful skill is knowing which direction to move it when retrieval disappoints.
| Symptom | Likely cause | Move |
|---|---|---|
| Exact identifiers, codes or product names are missed | Vector side dominating a query with little semantic content | Down — give keyword more weight |
| Right topic, wrong document; results feel adjacent | Keyword matching on a common term shared across documents | Up — let semantics discriminate |
| Nothing relevant returns for a plainly-worded question | Asker’s vocabulary differs from the author’s throughout | Up, and consider rewriting the document’s headings |
| A superseded document keeps winning | Not a weighting problem at all | Remove the document |
That last row matters more than the other three. Most retrieval complaints are document problems wearing a retrieval costume — a stale file, an ambiguous heading, a policy that exists in three versions. Tuning the weights on a bad corpus just changes which wrong answer you get.
The way to tell them apart is to run a test search rather than a conversation. Insulin’s test search shows a relevance score per result, so you can see whether the right passage came back at all. If it did and the answer was still wrong, the problem is downstream. If it did not, no amount of prompt-writing will save it.
What retrieval quality looks like in practice
A grounded answer in Insulin carries the source documents it drew on, and an agent can search up to three knowledge bases in a single turn. Two practical consequences:
Read the citations when you are evaluating, not the prose. The prose is generated and will be fluent regardless. The citations tell you whether retrieval worked, which is the thing you are actually assessing during a rollout.
Do not attach every base to every agent hoping for coverage. Three bases per turn means a broad attachment set spreads retrieval thin across material the agent did not need. Scope narrows the search space, and a narrower search space retrieves better — the same reason one base per domain beats one base for the company.
Frequently asked questions
What is hybrid search? Hybrid search runs vector similarity and keyword matching together and merges the results, so a passage surfaces whether the question used exact internal terms or described the idea in other words.
When does vector search fail? On strings with no semantic content — product codes, ticket numbers, clause references, internal acronyms. It returns something about the topic instead of the specifically named item.
Can I turn hybrid search off in Insulin? Yes. A knowledge base can be set to Hybrid, Vector only, or Keyword only. On hybrid, a vector weight between 0 and 1 sets how far results favour vector over keyword matching.
How do I tell a retrieval problem from a document problem? Run a test search rather than a conversation. It shows a relevance score per result, so you can see whether the right passage returned at all. If it did, the problem is downstream.
How many knowledge bases can an agent search at once? Up to three in a single turn. That is a reason to scope attachments narrowly rather than attaching every base to every agent in the hope of broader coverage.
Takeaways
- Keyword search is literal; vector search is semantic. Business documents break both individually.
- Hybrid runs both and merges, which is why named policies and vaguely-described ones both surface.
- Move the weighting down when exact identifiers are missed, up when results are topically adjacent.
- Most retrieval complaints are document problems — stale files and ambiguous headings — not weighting problems.
- Use a test search with its relevance scores to separate the two before tuning anything.
Insulin knowledge bases support hybrid, vector, and keyword retrieval with cited sources on every answer. Explore Insulin knowledge bases or get a demo.
Sources
Primary sources for the platform rules cited above. Last verified August 14, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure.
- Suger Insulin docs: Knowledge Base — Search mode is Hybrid (recommended), Vector only, or Keyword only; vector weight is how much to favour vector over keyword matches on a 0-1 scale; a test search returns matching 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.