Grounded by default
Every sentence carries a citation back to the exact chunk it came from. When the sources disagree, Lumen says so instead of inventing a tidy answer.
New
Lumen 2.4 — parallel tool calls & 400K context
Agents that read your data, use your tools, and show their work.
Lumen is a hosted runtime for production LLM agents. Connect your sources, describe your tools in a few lines, and ship an assistant your team can actually audit — with sub-second streaming and a bill you can forecast.
You
Review the Q3 vendor agreement and flag anything that deviates from our standard terms.
Lumen
search_documents(q: "vendor agreement Q3")
· 0.9s
compare(baseline: "MSA-standard-v4")
· 0.7s
I read all 84 pages and diffed them against your standard MSA. Three clauses deviate:
- §4.2 Payment terms are Net-60, not Net-30 — a 30-day cash-flow shift.
- §9.1 Liability cap sits at 1× fees; your policy floor is 2×.
- §12.4 Auto-renewal has no termination-for-convenience window.
I have drafted redlines for each. Want me to send them to Legal?
64 tok/s
210 ms to first token
Powering AI features at
Why Lumen
Prototypes are easy. Production is the hard part.
Anyone can wire a chat box to a model in an afternoon. Lumen handles everything that comes after — grounding, tool execution, evaluation, cost control and the audit trail your security review will ask for.
Tools in ten lines
Describe a function once. Lumen validates arguments against your schema, runs calls in parallel, retries on failure and streams the results back.
Streaming that feels instant
210 ms median time-to-first-token from 14 regions, with automatic failover between providers so a single upstream outage never becomes your outage.
Costs you can forecast
Per-request token accounting, hard budget caps per workspace and alerts that fire before the month runs away from you — not three weeks after.
Retrieval
Answers you can trace back to a paragraph.
Point Lumen at Notion, S3, Postgres, Google Drive or a plain folder of PDFs. It chunks, embeds and re-ranks on every write, so the index is never stale — and every response ships with the passages that produced it.
- Hybrid search — dense vectors plus BM25, re-ranked by a cross-encoder before the model ever sees a token.
- Row-level permissions — the retriever respects your existing ACLs, so nobody reads a document they could not open themselves.
- Freshness in seconds — incremental re-indexing on webhook, not a nightly batch job.
answer.py
from lumen import Lumen
client = Lumen(api_key=os.environ["LUMEN_KEY"])
# Ask a question against an indexed collection.
answer = client.answers.create(
collection="handbook",
question="How much parental leave in the EU?",
cite=True,
)
print(answer.text)
# → "EU staff receive 20 weeks at full pay …"
for c in answer.citations:
print(c.source, c.page, round(c.score, 3))
# → handbook.pdf 14 0.912210ms
Median time to first token, measured p50 across 14 regions
99.98%
Rolling 12-month uptime with automatic provider failover
6.4B
Tokens served every day across production workspaces
31%
Average token spend saved by semantic caching and routing
Use cases
One runtime, every internal workflow
Teams start with a single assistant and end up running a dozen. Same primitives, same billing, same audit log.
Support triage
Draft replies grounded in your macros and past resolutions, then hand off to a human with a one-line summary and a confidence score.
Knowledge search
Ask across Notion, Drive, Slack and the wiki at once. Permissions are inherited, so answers never leak past a team boundary.
Contract review
Diff incoming paper against your standard terms, flag deviations by clause, and produce redlines your counsel can accept or reject.
Analytics copilots
Translate a plain-English question into SQL against your warehouse, run it read-only, and explain the result with the query attached.
Code review
Comment on pull requests with repository context — conventions, prior decisions, and the ADR that explains why the odd bit is odd.
Compliance checks
Screen outbound copy, tickets and reports against policy before they ship, with an immutable record of every decision made.
How it works
Live in an afternoon, not a quarter.
No infrastructure to provision, no vector database to babysit. Three calls and your agent is answering real questions.
-
Connect a source
Point at a bucket, a database or a Notion workspace. Lumen chunks, embeds and keeps it fresh.
-
Declare your tools
Any HTTP endpoint or Python function becomes a callable tool with a validated schema.
-
Stream the answer
One endpoint returns tokens, tool events and citations over Server-Sent Events.
agent.ts
import { Lumen, tool } from "@lumen/sdk";
const lumen = new Lumen(process.env.LUMEN_API_KEY);
// 1 — a tool is just a typed function.
const refund = tool({
name: "issue_refund",
description: "Refund an order, up to €200.",
input: { orderId: "string", cents: "integer" },
run: async ({ orderId, cents }) =>
billing.refund(orderId, cents),
});
// 2 — stream tokens, tool events and citations.
const stream = await lumen.run({
model: "lumen-pro",
collections: ["support-macros"],
tools: [refund],
input: message,
});
for await (const event of stream) {
if (event.type === "token") write(event.text);
}Social proof
Shipped by teams who had to justify it
We replaced nine months of homegrown RAG plumbing with three Lumen endpoints. The part that sold our CTO was not the quality — it was the audit log.
Our support assistant handles 41% of tickets end to end. Every draft cites the macro it used, so agents trust it enough to hit send.
Token spend dropped 34% the week we turned on semantic caching, and I finally stopped getting surprise invoices on the first of the month.
Ship your first agent this week
Five million tokens a month, free forever. No credit card, no sales call, no six-week procurement dance before you can try it.
Average time from signup to first grounded answer: 11 minutes.