Technology
Infrastructure,
not interface.
Recora is built as cognitive infrastructure — a deterministic reasoning layer designed to operate under, inside, and alongside any workflow or system.
Sample Artifact Output
{
"artifact_id": "rec_a1b2c3d4",
"version": "1.0.2",
"domain": "real_estate.contract",
"source_doc": "purchase_agreement_2024.pdf",
"generated_at": "2024-11-01T14:32:00Z",
"confidence": 0.97,
"findings": [
{
"type": "MISSING_CLAUSE",
"severity": "HIGH",
"title": "No inspection contingency found",
"evidence": {
"quote": null,
"page": null,
"reasoning": "Standard purchase agreements require inspection contingency per §1234 state code"
}
},
{
"type": "RISK_FLAG",
"severity": "MEDIUM",
"title": "Unusual liquidated damages clause",
"evidence": {
"quote": "Buyer shall forfeit 10% of purchase price...",
"page": 4,
"section": "Section 8.2"
}
}
],
"citations": 12,
"hallucination_risk": "NONE"
}Every output is JSON-validated, versioned, and fully traceable to source. Hallucination risk is a first-class field.
Schema-First Design
Every domain in Recora has a formal schema that defines what can and cannot exist in an output artifact. Before any LLM transformation occurs, the IR is validated against the schema. Invalid data cannot propagate.
- ›Domain-specific JSON schemas for every use case
- ›Strict type enforcement on all extracted data
- ›Schema versioning with backward compatibility
- ›Custom schema extension for enterprise clients
Retrieval Architecture
Recora uses a hybrid retrieval approach — combining dense vector search with sparse keyword matching and structural position lookups. This ensures both semantic relevance and exact clause-level precision.
- ›Hybrid dense + sparse retrieval
- ›Hierarchical chunk position tracking
- ›Cross-document retrieval with source attribution
- ›Configurable retrieval depth per domain
Versioning System
Every reasoning run produces a versioned artifact. Artifacts can be compared, diffed, and audited. When a source document changes, downstream artifacts are flagged for re-validation — not silently invalidated.
- ›Immutable artifact versioning
- ›Diff engine for comparing artifact versions
- ›Source document change detection
- ›Full audit trail on all reasoning operations
LLM Abstraction Layer
Recora is architecturally independent of any specific LLM. The transformation layer accepts any model via a standardized interface. Switch from GPT-4 to Claude to a local model without changing your pipeline or output format.
- ›Model-agnostic transformation interface
- ›Prompt templates versioned alongside artifacts
- ›Multi-model ensemble support
- ›Local model support for sensitive data environments
Evidence Chain Validation
Every assertion in a Recora artifact must be traceable to a specific passage in the source document. The system enforces this at the generation layer — outputs that cannot be cited cannot be generated.
- ›Mandatory source citation for all claims
- ›Confidence scoring per evidence item
- ›Contradicting evidence detection
- ›"No evidence found" as a valid, explicit output
Domain Packs
Domain Packs are pre-built reasoning configurations that encode how experts in a specific field think. They include schemas, retrieval strategies, cognitive presets, and output templates — ready to deploy.
- ›Real Estate: contracts, disclosures, compliance
- ›Construction: SOW, specifications, codes
- ›Compliance: SOC2, NIST, ISO mapping
- ›Custom Domain Packs via professional services
From Probabilistic to Deterministic
The goal of Recora's architecture is to move AI-assisted reasoning from probabilistic outputs — where the system might be right — to deterministic reasoning systems where every output can be verified, audited, and trusted.
This isn't about making LLMs more accurate. It's about building systems that constrain LLMs to operate only on validated, structured data — so accuracy becomes a property of the architecture, not a hope about the model.