feature: SEO search-concepts/agentic search #105
Conversation
jamessewell
requested review from
philippemnoel and
rebasedming
as code owners
|
The latest updates on your projects. Learn more about Vercel for GitHub.
💡 Enable Vercel Agent with $100 free credit for automated AI reviews |
philippemnoel
left a comment
philippemnoel
left a comment
There was a problem hiding this comment.
I’m not sure about this one. I don’t really think we’re sharing anything super informative here. If you think it will help SEO, then sure.
I’m also not convinced that agentic search is different. Yeah the data is very structured and agents are good at writing SQL. But the whole iterating queries from results and such is exactly what humans do. Agents presumably do it faster, but is it really that different?
I’m also not convinced in the argument that it differs from RAG because RAG is hidden vs agentic search isn’t. At the end of the day people interact with agents via natural language and even though the chain of thought/set of search queries may be displayed as the agent thinks, it is still abstracted away.
What I think could be a more interesting topic to share about agentic search is that:
- agents are good at writing SQL, making search engines that are SQL native ideal. Eg nested queries with CTEs are hard for humans to reason about but easy for agents.
- Agents work by doing a lot of fast, short searches that build on each other. This makes keyword search a very strong fit for them, instead of vector search, which is inherently slower and more expensive (due to the vector search but also to the need to create embeddings). Agents search more « in real time ». Can use the Claude code announcements of using bm25 as an example
Those feel like more precise arguments to me that actually teach the user something new
| - **Programmatic Iteration**: Rapidly reformulate queries based on result analysis, testing different keywords, filters, or search methods within milliseconds | ||
| - **Persistent Context**: Maintain search context across multiple queries, using early findings to inform later query formulation and result evaluation | ||
|
|
||
| The most common way to enable agents is to expose search interfaces as tools in an MCP (model context protocol) server, but there are also other options like cetnering the search around a single technology that can access to all types of search. SQL makes a good candidate for this pattern. |
This one is a little wafty, but then so is the subject. It's also very, very poorly indexed so I would love to get this out ASAP