The integration keeps hitting rate caps
A chat endpoint that fans out one Sonar call per user turn, with no concurrency bucket or backoff, and a trial tier limit lower than the traffic you actually have.
Backoff in place, tier matched to load
Building on Perplexity?
Perplexity's search-grounded models look magical in a demo and start showing seams the moment you put them behind real traffic: citations that point at the wrong paragraph, a Sonar Pro call that cost ten times what you budgeted, or a response schema that quietly changed between model versions. We work the integration back into something predictable.
What it is
The Perplexity API, marketed as Sonar, exposes the same search-grounded language models that run the public Perplexity.ai product. Each completion runs a live web search behind the scenes, grounds the answer in retrieved sources, and returns an inline list of citation URLs alongside the generated text.
It ships Sonar, Sonar Pro, and reasoning variants. Pricing is a mix of per-million input and output tokens plus a per-request search fee, so costs behave differently from pure chat APIs. The REST surface is compatible with the OpenAI chat completions shape, which makes it easy to wire into an existing stack, as long as you remember the citation array and the fact that responses are not purely deterministic.
The honest version
No technology is all upside. This is where Perplexity API earns its place and where it tends to cause trouble.
Sound familiar?
The Perplexity API problems that bring people to us, and what you have once each one is fixed.
A chat endpoint that fans out one Sonar call per user turn, with no concurrency bucket or backoff, and a trial tier limit lower than the traffic you actually have.
Backoff in place, tier matched to load
A parser assuming citations are inline footnotes rather than a separate array, or a response where the model wrote more claims than it cited and the frontend silently dropped half of them.
Citations rendered accurately
An older online model identifier that Perplexity sunset without any code change on your side, surfaced as a plain 400 the morning after rollout.
Pinned to a supported model
A prompt that was moved from Sonar to Sonar Pro for a quality bump, combined with a UI that fires a search on every keystroke rather than on submit.
Spend back under control
An upstream search provider that is slow, no request timeout set in your HTTP client, and a worker pool that fills up waiting for responses that will never arrive.
Timeouts and retries that behave
Downstream code assuming every SSE event contains a citations array, when in reality the field only arrives on the final chunk for some model variants.
Stream handling hardened
A request passing search_domain_filter on a tier where the parameter is not honoured, so results still include the sites you explicitly tried to exclude.
Filtering working as documented
What we do
Bring us one failing endpoint, or hand over the whole Sonar integration.
How it works
Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.
We fix the cause, not the symptom. If something is down, getting it back up comes first.
Backups, monitoring, updates and access, so the same problem can't quietly come back.
You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.
Often found together
We work on the whole stack, so the problem can be wherever it actually is.
When the answer genuinely needs fresh web knowledge and you want the citations in the response rather than a separate retrieval step. For a product support bot grounded in your own docs, a retrieval layer over GPT or Claude will usually be cheaper and more controllable. For a research-style assistant that needs to quote current sources, Sonar is doing work that would otherwise be several tool calls.
The retrieval layer picks pages that look relevant to the query, and the model then writes a grounded answer, but it does not always point at the exact paragraph that supports a given sentence. If your product treats citations as proof, you need either a spot-check step or a second pass that re-ranks citations against the generated text.
Three things usually help. Cap requests per user session, route easy queries to the cheaper Sonar tier and reserve Pro for ones that failed quality checks, and debounce the frontend so it does not fire a search per keystroke. We also put a hard monthly ceiling at the API layer so a runaway loop cannot drain the account.
For non-critical uses, yes, with normal safety rails. For anything that must be accurate like financial, legal or medical content, we would pair it with a human-reviewed source list, a stricter domain filter, and a disclaimer on the output. The model hallucinates less than a pure LLM but it does still hallucinate.
Not in the same way. The API is focused on search-grounded chat, so if your use case needs structured tool calls or a complex agent loop, you will typically pair Sonar with another model that does tooling well, and use Sonar only for the knowledge retrieval step.
The estimate is free. Send over the endpoint, the current prompt and a sample of the responses that are misbehaving, and we return a price before any work starts.
Free estimate
Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.
Your message is with a real person, and we'll get back to you soon.
Rather talk now? Call us: