Is RAG dead? I would be very surprised a local small SOTA embedded model like llama-embed-nemotron-8b doesnt outperform the Haiku layer for this application. Should be pretty cheap and easy to prove out. With 32K context size, you can literally one shot the whole ticket.
We decreased our LLM costs with Opus
11–20 of 41 posts
Re: We decreased our LLM costs with Opus
#12> We switched to the "triager" pattern: a Haiku agent with a very specific and narrow job. Is this issue already tracked or not? If it is, stop right there. If not, escalate to Opus.
I'm planning to self host qwen3.6 27b basically for this purpose
Re: We decreased our LLM costs with Opus
#13Is RAG dead? I would be very surprised a local small SOTA embedded model like llama-embed-nemotron-8b doesnt outperform the Haiku layer for this application. Should be pretty cheap and easy to prove out. With 32K context size, you can literally one shot the whole ticket.
Yea, but RAG takes effort. At the very least some kind of system to organize the documents and do the retrieval.
My theory is that the AI frenzy has reached new levels of insane, where it's literally just throw anything and everything at the model, and just burn tokens to let the AI figure everything out. Why bother paying the upfront cost for a RAG, when the models/agents are constantly evolving, so just slap in a markdown file telling it to check a folder, and call it a day.
Like in design world, people are doing minor tweaks like changing the spacing by typing in prompts instead of just changing a number in an input field. We are legitimately approaching just using llms instead of calculators, or memes like that endpoint that calls an llm to generate the code to do some business logic, rather than directly code the logic.
Re: We decreased our LLM costs with Opus
#14> We switched to the "triager" pattern: a Haiku agent with a very specific and narrow job. Is this issue already tracked or not? If it is, stop right there. If not, escalate to Opus. > 4 out of 5 failures never reach Opus. A triager match costs around 25x less than a full investigation. The title feels misleading. Why clickbait on that when you can just be genuine about the architecture?
I am one of Mendral co-founder (my co-founder wrote the article), I am the one to blame for changing the title when posting. I thought our original one was too clickbait and I wanted to better summarize with this title.
Despite the original title, a lot of what we learned comes to how Opus evolved and the ability to reason. And also the fact that Haiku is quite capable if scoped properly, that's the whole purpose of the article.
Re: We decreased our LLM costs with Opus
#15I have rewritten the article to be slightly shorter: “Let a cheap agent decide if the expensive one is needed.”
Sounds like L1 vs L2 support :)
It's the same as an escalation. Something we omitted from the post is that we often use Sonnet to write SQL queries.
We wrote another post that was on HN some time ago that goes into the details of SQL queries (linked at the top of this article). Sonnet is perfect for this.
Re: We decreased our LLM costs with Opus
#16Looking at the diagram, is this seriously a case of replacing basic functional concepts like "write to clickhouse" or "have we seen this before" to a model? could those be actual function calls in some language? just seems wasteful all around. having an agent in the critical path when a regular expression (or similar) could do just seems odd. yeah haiku is cheap but re.match() is cheaper.
We're dealing with CI logs, produced by a variety of frameworks, languages, etc... And the tough ones to look into are e2e tests, with outputs from infrastructure.
I wish a re.match() would be enough, but we often don't even know what to match in the first place.
We started to add deterministic matching on the patterns that the agent sees the most so we don't have to go through the whole thing (for example a flake on PostHog can occurs 100+ times during a day, you don't need to reinvestigate every time). But for new errors, it's tricky.
Re: We decreased our LLM costs with Opus
#17Is RAG dead? I would be very surprised a local small SOTA embedded model like llama-embed-nemotron-8b doesnt outperform the Haiku layer for this application. Should be pretty cheap and easy to prove out. With 32K context size, you can literally one shot the whole ticket.
IMO RAG is mostly dead. The game changer with newer models like Opus is the reasoning. So instead of pushing all the context up front (RAG style), it's better to give strong primitives (eg. bash, SQL) and let the agent figure it out.
It's what Claude Code is doing now and the principles we applied for Mendral as well.
That said, you're right that some smaller models can outperform Haiku and we're thinking supporting oss models at some point. But it does not change the core design principles IMO.
Re: We decreased our LLM costs with Opus
#18I do a similar thing with a "planner agent" that uses the cheapest (I think it's using openai-gpt-5.2-mini or something at like 20 cents for 1M.) that more or less emits a plan name, task list and the task list has a recommended model in each task. It's not perfect, but many of our tasks are accomplished with lighter weight models. When doing code generation or fixing we upgrade to a more expensive model, planning an…
Curious, what steps did you follow to end up with this design (what did you try before)? And what's your use case for this agent?
Re: We decreased our LLM costs with Opus
#19[dead]
Re: We decreased our LLM costs with Opus
#20I want to create a "harness" that does this with Claude Code and other expensive agents.
Buffer user prompts, use conversation history and repo state as context -- and run a local model or a cheap and fast cloud model like Haiku to determine the optimal way to address the user's ask, reframe the query with better context (user reviews and approves if needed) and THEN let expensive models like Opus have a go at it.
If we are operating within Anthropic ecosystem with Haiku and Opus -- this sort of logic should ideally be doable within Claude Code as harness. Currently skills cannot be tagged to different models. Ideally we should be able to say -- for trivial tasks, the skill should always use Haiku even if invoked from a session with Opus xhigh.