SuperChargeDBBy ZadeNor AIThe search engine for the multimodal era.
An object-storage-native, multimodal vector + document engine that returns the exact moment, region, passage, or line — not just a document id. Hybrid retrieval, cited agentic answers, and enterprise-grade access control, built on a modern serverless cloud platform.
Multimodal planes + Locator
Four planes fuse into one answer that points to the exact spot.
Most vector databases stop at the row. SuperChargeDB indexes image, audio, video, and document modalities across four planes — text, visual, audio, and doc-visual — and returns a fine-grained Locator (the precise moment, region, passage, or line), fused by Reciprocal Rank Fusion.
Text
→ passage · line span
Documents, transcripts, and code chunked, embedded, and indexed for dense + lexical recall.
Visual
→ frame · bbox region
Images and video frames localized to the exact moment and bounding-box region.
Audio
→ timestamp window
Speech and sound aligned to precise timestamp windows across long-form media.
Doc-visual
→ page · figure · cell
Scanned pages, tables, and figures via MUVERA visual-doc handling — down to the cell.
Hybrid + agentic search
Answers you can trust — every claim cited, every claim verified.
Ask in natural language. A single POST /v1/answer runs a retrieve → localize → verify → cited-synthesis loop with NLI faithfulness — grounding every claim in the exact moment, region, passage, or line it came from.
The live failover is demonstrated 12:47 into the keynote, where the primary region is drained and traffic re-homes with no dropped requests[1]. The runbook confirms a sub-second cutover[2].
Retrieve
Cheap, high-recall hybrid search across every plane — dense vectors + BM25 fused by RRF.
Localize
Narrow to a tiny top-k and pin the exact Locator: moment, region, passage, or line.
Verify
An NLI faithfulness pass checks every claim against its source before it's allowed to speak.
Cite
Synthesize a grounded answer where every sentence carries a verifiable citation.
Built for every domain
One retrieval core. Every industry's hardest questions.
From patient scans to case law to keynote footage — SuperChargeDB returns the exact moment, region, passage, or line, with cited answers and enterprise-grade access control tuned to your domain.
Healthcare
Retrieve the exact scan region, chart passage, or clinical note — every hit clamped to patient- and role-scoped access.
Legal
Pin the precise clause, page, and citation across contracts and case law — cited, verified, and audit-ready.
Finance
Search filings, transcripts, and dashboards down to the line — grounded answers with a tamper-evident trail.
Media & Entertainment
Find the exact frame, moment, and quote across video, audio, and archives in seconds — not hours.
Government & Defense
DoD-ready retrieval with hash-chained WORM audit and per-principal scope enforcement across every surface.
Retail & E-commerce
Unify catalog, imagery, and support content into one multimodal, shoppable search surface.
Enterprise access control
One enforcement core. DoD- and enterprise-ready.
Cross-tenant, cross-project, or cross-principal leakage requires multiple independent failures. Every hit is filtered by acl_owner / acl_groups at the ANN metadata layer before it ever reaches you.
Scopes below namespaces
A first-class project dimension nested under each tenant namespace — two independent isolation levels, zero per-tenant infrastructure.
Per-user ACL narrowing
Every request is clamped server-side to effective_scope = requested ∩ granted. The caller — or the LLM — can never widen access.
Full RBAC / ABAC
Role- and attribute-based grants enforced by one core across every surface, backed by API-token → principal resolution.
Verifiable hash-chained audit
Tamper-evident audit sealed into object-lock (WORM) storage; the chain head is CAS-committed and independently verifiable via GET /v1/audit/verify — any reader can prove the log was never reordered.
Five surfaces, one core
Reach the engine any way you build. Same enforcement everywhere.
There's no privileged surface and no path to your data that skips the clamp. The TypeScript + Python SDKs, the CLI, and the remote MCP server all ship today — with identical RBAC/ABAC and scope intersection across all five.
UI
Access / SSO session
CLI
SUPERCHARGEDB_TOKEN
REST
Bearer / Access-JWT
SDK
TS + Python client
MCP
OAuth 2.1 · agent-native
// Same query, same scope clamp — UI, CLI, REST, SDK, or MCP.
const { hits, answer } = await turbo.search({
query: "failover during the Q3 keynote",
scope: "acme/prod/keynotes", // intersected with your grants
planes: ["text", "visual", "audio", "doc_visual"],
fusion: { rrf: 60 }, rerank: true,
cite: true, // verified, grounded answer
});
// → hits[0].locator = { t: 767.3, plane: "audio" }Cloud-native architecture
We stopped building the two hardest pieces. The platform hands them to you.
No storage engine to run, no ANN index to tune. The scarce resource flips from object round-trips to inference compute — so the rule becomes: retrieve cheaply, spend deliberately.
Object storage
Egress-free source of truth
Vector index
Managed ANN index across every plane
Managed inference
Embed · rerank · VLM · ASR (usage-billed)
Coordination
Manifests, sessions, strong-consistency delta
Queues + Workflows
Durable ingest, compaction, reconcile
Containers
CPU-bound ffmpeg · OCR · scene detection
Agent tooling / MCP
Agent-native remote tooling
Inference gateway
Cache · rate-limit · per-model cost & latency
Napkin math
50k documents · 200k queries / month
Estimated total
~ $80–90 /mo
4
Multimodal planes
text · visual · audio · doc-visual
5
Access surfaces
UI · CLI · REST · SDK · MCP
k=60
RRF hybrid fusion
dense + BM25, reranked
$0
Egress fees
object storage source of truth
Ship retrieval that points to the answer.
Run the exact same code locally and on the global edge. Local → prod parity, zero rewrite.