The wrong passage reaches the model.
Forge provides the embedding layer; Answers returns ranked evidence and a cited result.
KNOWLEDGE IS PRECIOUS.
Call one managed API from an agent, or publish the same corpus as a private answer site. Voxell handles the retrieval pipeline and shows you where the corpus is weak.
FREE TO START · NO CARD · PRODUCTION USE ALLOWED Requests pause at the limit. Nothing bills without you choosing it. What’s included →
POST /v1/answers/employee-policy/query
$ “Can annual plans be refunded after renewal?”
200 retrieval.health=healthy score=0.91
“Yes. Annual plans may be refunded within 30 days of the renewal date.”
[1] refunds.md §4.2
ANSWER + RANKED EVIDENCE RETURNED
CITED ANSWER — THE WHOLE PIPELINE IN ONE CALL
curl https://api.voxell.ai/v1/answers/employee-policy/query \
-H "Authorization: Bearer $VOXELL_KEY" \
-d '{ "query": "Can annual plans be refunded after renewal?", "mode": "answer" }'
RAW EMBEDDING — DROP-IN FOR THE OPENAI SDK
curl https://api.voxell.ai/v1/embeddings \
-H "Authorization: Bearer $VOXELL_KEY" \
-d '{ "model": "text-embedding-3-small", "input": "Can annual plans be refunded after renewal?" }'
Swap mode to vectors for passages without the answer. Answers reference → · Integrations →
kb.acme.comPRIVATE
“Can annual plans be refunded after renewal?”
Yes. Annual plans may be refunded within 30 days of the renewal date.
[1] refunds.md §4.2
Numbers, with receipts: 87 ms is P50 end to end on owned NVIDIA DGX hardware, no shared queue. Ingot-8B-R3 scores 75.98 Mean (Task) on English MTEB v2, rank #1 as of July 2026, independently verified on the public MTEB leaderboard.
THE FAILURE STARTS UPSTREAM
A better generator cannot recover the passage your retriever missed, a broken chunk, or two sources that disagree. Voxell makes that path visible and replaceable.
Forge provides the embedding layer; Answers returns ranked evidence and a cited result.
VQS flags split damage, near duplicates, junk attractors, and contradictions before users do.
Chunking, indexing, hybrid weights, reranking, context assembly, citations, and evals become one call.
ONE PLATFORM // THREE DISTINCT JOBS
YOU HAVE A RAG STACK
Swap the OpenAI base URL. Keep your vector database, orchestration, and generator.
YOU HAVE DOCUMENTS
Upload a document, ask a question, receive grounded JSON with the source passages attached.
YOU HAVE AN AUDIENCE
Publish a corpus as a private answer site with access controls and an answer budget.
THE PIPELINE TAX
Keep control when you need it. Delete the assembly line when you do not.
READ THE REAL API SHAPEPOST /v1/answers/{corpus}/query
vendor-contracts
214 documents · 8,912 chunks
VECTOR QUALITY STANDARD // VQS
VQS scores the corpus itself before anyone asks a question. It shows whether the material is the bottleneck—or whether retrieval is.
ZERO TRUST PLATFORM ARCHITECTURE // STATED PRECISELY
Separate the stateless embedding path from products that persist a corpus. Scope authentication and retention claims to the deployment that provides them.
PROOF BEFORE PROCUREMENT
Start self-serve, or send 10–50 failed queries and a sanitized corpus slice. Get retrieved passages and a fit / no-fit result—no intro call required.