Hybrid
Dense + sparse retrieval
Vector similarity finds semantic matches even when wording differs. Keyword search catches exact identifiers, codes, and quoted phrases. Both run in parallel; results merge.
caveauAI Architecture
Under the hood: a corpus you can inspect.
caveauAI is a multi-stage retrieval architecture, not a chat wrapper. Eight stages between document and cited answer, three deployment topologies, and a model layer you can swap. Built for engineers and CTOs evaluating private RAG seriously.
The pipeline
Eight stages from document to cited answer.
Each stage is independently observable, tunable, and replaceable. Stage boundaries are the points where compliance, audit, and customization happen.
01
Ingest
Documents enter the pipeline in raw form — PDFs, scanned images, audio recordings, plain text, or structured exports from email and SharePoint. Audio is routed through GPU-accelerated speech-to-text producing timestamped, speaker-segmented transcripts.
02
Normalize
OCR for scanned material, structural parsing for native files, language detection, metadata extraction (author, date, document type). Output is a clean, machine-readable document object with provenance intact.
03
Chunk
Documents are split at semantic boundaries — paragraphs, clauses, sections — not arbitrary character counts. Chunk size and overlap are tunable per corpus. Each chunk inherits parent-document metadata.
04
Embed
Each chunk is encoded as a high-dimensional dense vector via a domain-appropriate embedding model. Embeddings capture semantic meaning, not keywords — enabling retrieval that survives paraphrase, translation, and synonym drift.
05
Index
Vectors land in a per-tenant vector store, partitioned by corpus. Sparse keyword indices are maintained in parallel for hybrid retrieval. Metadata filters (date, author, document type) attach at index time.
06
Retrieve
A query is embedded with the same model, then run as a hybrid search: dense vector similarity plus sparse keyword match. Metadata filters narrow scope. Top-k candidate chunks come back with similarity scores.
07
Rank
Candidates are re-ranked by a cross-encoder model that scores query-chunk pairs directly. This second pass corrects for embedding-space limitations and surfaces the chunks that actually answer the question.
08
Cite
A language model generates the answer grounded in the ranked chunks, with explicit citation back to the source passage. The user sees the answer and the evidence side by side.
Retrieval strategy
Hybrid, filtered, re-ranked.
Pure vector search is fragile on long-tail queries. caveauAI combines three retrieval mechanisms to surface evidence that actually answers the question.
Filtered
Metadata-aware narrowing
Filter by date, author, document type, jurisdiction, client, or any custom field captured at ingest. Reduces noise on large corpora.
Re-ranked
Cross-encoder second pass
After initial retrieval, a smaller cross-encoder model scores query-chunk pairs to surface chunks that genuinely answer the question, not just chunks that match in vector space.
Model layer
No single-vendor lock-in.
caveauAI is provider-agnostic at the model layer. Embeddings, retrieval re-ranker, and generation can each be swapped without rebuilding the corpus.
Open-weight
Self-hosted open models
Run on Blue Note Logic GPU infrastructure or your own. Common choices: Llama, Mistral, Qwen — selected per corpus characteristics.
Hosted API
OpenAI, Anthropic, Azure OpenAI
Routed via a unified gateway with key management, rate limiting, and request logging. Per-tenant policy on which providers are permitted.
BYO key
Your provider, your contract
Bring your own API key from any supported provider. Useful when you have an existing enterprise agreement or compliance posture.
Fine-tuning
Path to custom models
When retrieval alone is no longer enough, Blue Note Logic can scope fine-tuning on a customer corpus. Available on Sovereign deployments.
Deployment topology
Three boundaries. One workspace.
The application, API surface, and tool behavior are identical across topologies. What changes is where the data sits and who operates the infrastructure.
SaaS
Shared multi-tenant
Fastest path to running. Blue Note Logic operates the full stack in EU bare-metal datacenters with per-tenant corpus isolation.
- EU datacenter (Hetzner / equivalent)
- Per-tenant vector store partition
- Shared application tier
- Standard SLA
- Best for: pilots, SMB, advocacy
Private Cloud
Dedicated single-tenant
Single-tenant deployment on dedicated infrastructure — Blue Note Logic bare-metal or your VPC (AWS, Azure, GCP).
- Dedicated compute, dedicated storage
- Customer-controlled retention policy
- Custom embedding/model selection
- Enhanced SLA + dedicated support
- Best for: mid-market, regulated industries
Sovereign
On-prem / air-gap
Full caveauAI stack installed inside your perimeter. No outbound dependency on Blue Note Logic infrastructure or public model APIs.
- Air-gap capable
- Open-weight models only (or BYO)
- Customer-operated or BNL-managed
- Fine-tuning supported
- Best for: government, defense, critical infrastructure
API surface
Three integration paths.
Detailed API documentation lives on the all-features page. Short version below.
REST API
Query, chat, document CRUD
Standard REST endpoints for corpus management, retrieval, chat completion, and document lifecycle. API key auth with scoped permissions.
MCP
Model Context Protocol
Expose caveauAI corpora to MCP-compatible agents (Claude Desktop, Cursor, custom agents). Grounds external assistants in your private knowledge.
Workflows
Webhooks + scheduled jobs
Agent Builder chains retrieval, analysis, and delivery into repeatable pipelines. Trigger on incoming email, file drop, or schedule.
Need to see this against your stack?
Blue Note Logic engineers will walk through caveauAI architecture against your existing data infrastructure, identity model, and compliance posture.