Three layers of intelligence.
Each one makes the next more powerful.

Adaptive document understanding

Tessera doesn't treat every document the same. Semantic chunking preserves the logical structure of legal clauses. RAPTOR creates hierarchical summaries so both fine-grained clauses and document-level themes are searchable. Structural parsing handles headers, tables, and nested lists that break simpler systems.

Hybrid retrieval with intelligent fusion

Every query runs against both a vector index (semantic similarity) and a knowledge graph (entity relationships). Results from both paths are fused using configurable strategies — weighted scoring, reciprocal rank fusion, or Borda count — then re-ranked by a cross-encoder that evaluates each result against the original query for precision.

Schema that emerges from content type

A collection of supplier agreements produces a different metadata schema than a collection of research papers. Tessera detects document type, extracts type-appropriate metadata, and adapts its retrieval strategy accordingly — before any human configures anything.

Entities and relationships at scale

As documents flow into Tessera, every company, person, clause type, jurisdiction, risk factor, and concept is extracted and mapped into a knowledge graph. Relationships carry types and weights — 'COMPETES_WITH' behaves differently than 'SUPPLIES_TO' in query traversal.

Cross-document reasoning

The graph enables questions that span documents: Which counterparties appear across multiple contract families? How has a specific entity's liability language changed across contract vintages? What downstream dependencies exist if a supplier relationship is terminated? These questions are unanswerable by searching individual documents.

Temporal relationships

Relationships have time dimensions. Tessera tracks when entities first appear, how their connections evolve, and when relationships change character. This enables temporal queries: 'Which risk factors appeared for the first time after 2020?' — answered not by document dates, but by relationship emergence in the graph.

The central innovation: emergent metadata

Traditional metadata systems require someone to define what's important before the system can find it. Tessera inverts this: AI agents analyze the corpus and discover what's semantically significant. The schema materializes from the content itself. Different corpora produce different schemas — because different domains have different things that matter.

Cross-variable correlation at portfolio scale

Many of Tessera's most powerful findings come from correlating variables that no human would think to cross-reference. The Grace Period Trap (54% of grace periods have built-in nullification clauses) required correlating three independent variables across 1,910 documents. The Bank vs. Credit Union Chasm required inventing a dimension (issuer type) that wasn't in the data, extracting it from raw text, then discovering it was the strongest predictor in the corpus.

Zero-configuration domain adaptation

Tessera analyzed 510 commercial contracts and found indemnification gaps and IP assignment blind spots. Then it analyzed 1,910 consumer credit card agreements — with zero configuration change — and found arbitration traps and grace period nullification. Same engine. Different domain. Different findings. The intelligence is emergent, not engineered.

Term Discovery

Reads every document and extracts thematically significant terms — the specific ideas that matter, not the boilerplate that surrounds them.

In commercial contracts, this might surface 'uncapped indemnification' or 'change-of-control gap.' In credit card agreements, it surfaces 'mandatory arbitration' or 'penalty rate trigger.' The engine doesn't know the domain — it discovers what's distinctive.

Consolidation

Clusters terms across the entire corpus. Variants of the same idea — however differently expressed across documents or authors — resolve into a single canonical concept.

'Limitation of liability,' 'liability cap,' and 'maximum aggregate liability' become one concept. This normalization is what makes cross-document analysis possible at scale.

Pattern Analysis

Analyzes each concept across time, source, geography, and co-occurrence. Is this theme growing? Concentrated? Spreading?

This wave found that arbitration clauses tripled after 2018 while penalty rates declined — a strategic enforcement shift that's invisible in any individual contract.

Quality Review

Reviews the discovered schema for quality and coherence. Noise is filtered. Only meaningful intelligence survives.

A reflection and critique loop validates each discovered pattern. Low-confidence findings are discarded. What remains is intelligence Tessera is willing to stand behind — with citations.