AI Search Has a Problem – It’s the Answering Part
For Better AI Search: Clarify First. Search Second. Answer Never
Most AI-powered search tools have a common flaw. They go for the answer.
They look great in demos, but fall apart in real-world use.
Traditional search had a simpler failure mode. It matched keywords, ranked pages, and handed you a stack of links. Crude, but honest. The machine was not pretending to know. It was showing you what it found.
AI search changed the surface, not the risk. Now the system reads a vague question, guesses what the user meant, blends material from different sources, and returns a polished answer with the tone of authority. Sometimes it is right. Sometimes it is confidently wrong. Knowing the difference? That is the whole problem.
ClaRA came out of a SaaS support issue. Users had questions. The answers existed in the knowledge base. Support still got the calls. The problem was never just missing content. It was failed retrieval. The articles used the right terminology for keywords that the users did not include in their queries. Support staff knew how to translate a confused question into the right search terms. Regular users did not.
At first, a standard RAG approach looked like the obvious fix: limit the AI to the knowledge base and let it answer from there. In practice, that still left too much room for interpretation. The model would read the material, infer an answer, and occasionally get it wrong anyway.
So we stopped trying to make the model an authority. Instead, we built ClaRA.
ClaRA stands for Clarifying Retrieval Agent, and the name is the design decision: It clarifies first. Then it retrieves.
The Big Insight
ClaRA uses two distinct agentic workflows because there are two different jobs.
Agent A
Agent A handles clarification.
Its job is not to answer the user’s question. Its job is to make the question answerable. In three turns or fewer, it asks for the missing context: the form, the year, the jurisdiction, the screen, the workflow, the error code, whatever closes the gap between “something is wrong” and “here is the actual issue.” Then it rewrites the request into a clean structured query and passes it forward.
Agent B
Agent B handles retrieval.
It does not summarize its way into trouble. It takes the clarified query and returns the most relevant knowledge base articles in ranked order.
The user decides which articles will provide the information needed.
The agentic separation matters because these are different problems. Clarification is about intent. Retrieval is about search. Splitting the problems allows each agent to improve on its own. That was the bet behind ClaRA. It turned out to be the right one. (This is a little bit like Coyote Time but for search)
Isn’t this an extra step?
Which is more frustrating? A clarifying question, or getting bad results?
Most users do not ask well-formed software questions. People ask from a confused state. They describe symptoms instead of causes. They omit the year, the exact form, the screen name, or the step where things went sideways. They assume the system sees what they see.
The search box does not know what the user wants, it only knows what the user has typed.
ClaRA slows down at the only point where slowing down helps.
Not at the end. At the beginning.
A search tool can only retrieve what the question makes retrievable. When the input is vague, the output usually is too. That is why so many self-service experiences fail before the search engine even starts. The system is trying to retrieve against a question that is still half missing.
ClaRA fixes that by slowing down at the only point where slowing down helps.
Not at the end. At the beginning.
A small amount of clarification up front produces better article matches, fewer dead-end searches, and fewer support calls from users who already tried the knowledge base and got nowhere.
That is the real self-service win. Not forcing users away from calling support. Making the help system useful enough that they do not need to call support in the first place.
Why ClaRA returns links instead of answers
ClaRA returns ranked article links instead of AI-generated final answers on purpose. It’s a choice that runs against the current trend, which is exactly why it matters.
While plenty of companies are eager to provide AI results ahead of authoritative links (looking at you Google), the results are not good enough, especially when the answers are used for business-critical decisions. Trust me, if you’re costing your customers, it’s costing you customers.
LLMs are strong at understanding natural language, spotting ambiguity, and helping narrow a question. They are less reliable as the final authority in domains where precision matters. They blend contexts. They pull the wrong fact from the wrong place and deliver it like it came down from the mountain on stone tablets.
It’s funny when it’s trivial, but it’s dangerous (and expensive) in other contexts
ClaRA uses AI for the part AI actually does well: understanding language well enough to improve retrieval. Then it stops.
That restraint is the feature.
The user gets the source material. The support rep gets the source material. The source material is still the authority. The AI just helps you find it.
The takeaway
A lot of AI search tools rush to sound useful.
ClaRA works because it follows a simpler sequence:
- The user asks a messy question.
- Agent A clarifies it.
- Agent B retrieves the right sources.
- The human applies judgment.
That makes support faster and it makes self-service more trustworthy.
And it solves the actual problem, which was never “how do we generate more answers?” It was “how do we help people find the right one?”
