@hwellmake Hello! Great question.
Short version: Yes, Katzilla has natural language API calls that like in multiple government sources in a single application – a credit. But that’s on a different endpoint/different product. Your question makes a good point and an opportunity for improvement. By the time you see this, our main product endpoint /v1/nl will route to multiple actions when the prompt means it!
1. /v1/nl Now is multi-source.
Previously, our NL endpoint chose exactly one best match action per call. You were right – it was meant to be a question. “FDA pulls back on pharma companies with recent SEC filings”. will either take the wrong path or force the agent to make N separate calls on N credits.
Now, Claude’s router returns. 1–3 routes per call and we fan them out in parallel. A credit, multi-source result:
POST /v1/nl{ “query”: “Compare US inflation to unemployment trends in 2024”, “execution”: true }
→ {
Routings🙁
{ Agent: “economic” Action ID: “Fred Series”, Params: { Series ID: “CPIAUCSL” } },
{ Agent: “economic” Action ID: “Fred Series”, Params: { Series ID: “UNRATE” } }
)
Data: {
“economic/fred-series”: { … }, // keyed dict, easy to index
…
},
Results🙁 … ), // per source description + references
Meta: { Sources_inquiry:2, Sources_success: 2 }
}
Single-source queries still return the legacy routing field, so existing integrations are not broken.
2. /v1/ask/answer – Synthesized cross-source response
For cases where the agent should not perform client-side synthesis, we’ve added a new endpoint that takes the raw parallel results from the router and runs the final cloud pass to generate. A prose response with inline (1)/(2) References:
POST /v1/ask/answer{ “q”: “Memoranda and Recent 10-Q Disclosures of Pharma Firms with Class I FDA”, “max_sources”: 3 }
→ {
Answer: “Three firms have posted Class I recalls since January (1), and two of them flagged related product liability language in their most recent 10-Q (2)…”
References:({ I.D:1, Source: “Open FDA”, Reference: {…} }, …)
Results🙁 … )
}
Both endpoints charge. One credit per call regardless of fanout — Same billing model as /v1/ask. on the documentation page katzilla.dev/docs Now there is a level guide so agents choose the right endpoint:
/v1/nl — Plain English, single or cross-source, cloud roots
/v1/ask — keyword-rooted multisource (cheap, no LLM from us)
/v1/ask/answer — multisource + synthesized prose answer
/v1/compose — You already know the sources.
Thanks for your question, welcome!