The maintained knowledge layer
- Approved external sources and immutable versions
- Canonicalization, chunking, retrieval profiles, and updates
- Source lineage, citations, evaluation, activation, and rollback
RAG as a service for maintained knowledge
Give your AI knowledge it can trust.
CorpusMesh is a managed RAG service for authoritative knowledge that must stay current. We maintain source boundaries, versions, retrieval indexes and citations while your application keeps control of its model and user experience.
- Versioned sources
- RAG with citations
- Your model stays yours
Private beta. Access is reviewed manually and public retrieval is not available.
Live lineagePortfolio mechanism · source to citation
Official sourceApproved authorityidentity · rights · source scope
Source evidenceCanonical recordURL · rights · checksum
Change signalAuthority monitoredreview before activation
Immutable versionCandidate v01chunking + retrieval profile sealed
query + allowed filters
Ranked passageResult 01 boundedboundary and citation checked
Ranked passageResult 02same authorized version
Citation attachedAuthority · section referencesource identity survives retrieval
Source operationsSource-specific monitoringreview before activation
Corpus boundaryBounded for one workflowsize and update cost approved first
Evidence returnedRanked passages + citationsversion and source metadata
Customer controlYour model and UX stay yoursone named application
Private-beta portfolio
Private betas for high-value knowledge workflows.
Every listed knowledge base is available through our private-beta programme, with coverage scoped to your product, jurisdictions, and update requirements.
Featured private beta01
Financial services regulation
Traceable regulatory material for financial products operating across US and European markets.
- Configured coverage
- US and European regulatory coverage configured around your regulators, products, and compliance workflow.
Education private beta02
Education curricula
27,015 indexed curriculum points, with a sealed 108-query Drafted evaluation reporting zero cross-boundary hits.
- Configured coverage
- Official curriculum coverage for France, California, Florida, New York, Texas, England, Scotland, Wales, Northern Ireland, Poland, and Mexico.
All private betas03
Six high-value knowledge bases
Finance, trade, procurement, filings, regulation, and education, qualified around one approved workflow.
- Portfolio boundary
- Six knowledge bases · three delivery routes · one approved workflow at a time
Knowledge from the sources your AI needs to cite.
Maintained, versioned, attributableOfficial sourcesCurricula, legislation, registries
Open datasetsPublic data, research, standards
Licensed collectionsExpert-maintained knowledge
Where CorpusMesh fits
Maintained external knowledge, without owning the retrieval pipeline.
CorpusMesh handles the source and retrieval layer for bounded authoritative domains. It does not replace the product, model, or judgment your team owns.
Your application and its decisions
- No generic internal-document chatbot
- No generated legal advice or automated compliance decision
- No lock-in to a CorpusMesh model, prompt, agent, or interface
Service boundary
Internal document RAG and maintained external knowledge solve different jobs.
| Responsibility | Internal document RAG | CorpusMesh |
|---|---|---|
| Primary sources | Your organization's files and uploads | Approved authoritative external sources |
| Updates | Your team detects changes and reindexes | CorpusMesh monitors, reviews, versions, and activates |
| Retrieval evidence | Depends on your implementation | Source anchors, versions, checksums, and evaluation |
| Answer generation | Your application | Your application |
Published evaluation · 80 fixed queriesMeasured retrieval, with its limits visible.
This is one evaluated EU AI Act workload, not a universal quality claim.
- Recall@10
- 96.2%
- nDCG@10
- 82.0%
- Boundary violations
- 0
Two ways to use CorpusMesh
Build with maintained knowledge. Or bring knowledge worth maintaining.
The same versioned infrastructure connects teams building AI products with the sources those products need.
The knowledge layer
One maintained layer between trusted sources and your AI.
CorpusMesh owns the difficult pipeline so your team can focus on prompts, agents and product behavior.
One verified implementation
Built to remain inspectable.
The EU AI Act proves the shared factory: collection, provenance, structure, retrieval evaluation, activation, and rollback.
EU AI Act referenceINTERNAL VERIFIED · 2026-08-24.3
105sealed fixtures
0boundary violations
2,298active chunks
Retrieval profiles
Choose quality, not infrastructure lock-in.
Dense, lexical, hybrid and reranked retrieval can be versioned per knowledge base.
Access control
The right context, for the right client.
Entitlements, filters, bounded responses and exposure monitoring are part of delivery.
Published lineagereviewed v1 artifact
Source recordEUR-LexRights evidenceReuse recordedSource metadataAuthority sealed
Immutable version2026-08-21.1checksum + retrieval profile sealed
Ranked passageResult 01 0.850Article 6 Classification rules for high-risk AI systems
Your applicationbounded context + citation
Source trace
Every passage keeps its history.
From the original document to the exact retrieved chunk, each result carries its authority, version, checksum and citation path. The published benchmark and trace remain pinned to the reviewed v1 artifact; v2 is active internally and has not been authorized as a public trace.
- Immutable source versionsIncluded
- Deterministic lineageIncluded
- Atomic activation and rollbackIncluded
Built for your stack
One request. Context your AI can use.
Send a raw query and structured filters. The prepared private-beta REST boundary returns ranked passages, source metadata and citations.
Your LLMstays under your control
Portableacross models and agents
{
"query": "How do Article 6 and Annex III work together to identify high-risk AI systems?",
"knowledge_base": "eu-ai-act-reference",
"filters": { "jurisdiction": ["eu"] },
"top_k": 5
}200 OK5 passages · 5 citations · reviewed trace
CorpusMesh