On ingest, Primer standardizes metadata and identifiers, then parses text into entities, claims (derived statements or facts), and semantic chunks with source spans. These features are the building blocks that power each phase of the downstream analysis.

At query time, they deliver precision, speed, and traceability.

For example, the sentence:
“Several unauthorized drone incidents occurred along the U.S.-Mexico border between May 15 and May 30, 2025, involving models such as Dragonfish, Skydio 2+, Black Hornet Nano, and others.” produces the following claims:

• Several unauthorized drone incidents occurred along the U.S.-Mexico border between May 15 and May 30, 2025.
• The incidents involved models such as Dragonfish.
• The incidents involved models such as Skydio 2+.
• The incidents involved models such as Black Hornet Nano.
• The incidents involved other drone models.

The Primer platform uses a pretrained transformer model to identify more than 20 entity types in text with precise character-level source spans that let users validate results against exact snippets of text. These extractions enable fast entity-based filtering, search, and aggregation across the entire corpus without reprocessing.

Entity types: people, organizations, locations, weapons, vehicles, dates, currencies, diseases, devices, animals, quotes, and more

A step beyond, NER, resolve the mentions of an entity across documents using a multi-tiered contextual matching system.

For people and organization entity types, the system progressively applies document-level alias detection and contextual analysis, with a fine-tuned language model for resolving ambiguous cases.

This links variations like "FBI", "Federal Bureau of Investigation", and "the Bureau" to a single canonical entity in our knowledge base. For all other entity types (weapons, vehicles, and others), the system uses a disambiguation algorithm to group similar mentions, linking "MQ-9 Reaper" with "Reaper drone" across documents.

This multi-stage resolution lets users generate relationship networks and unified entity profiles for people and organizations, while ensuring that entities are consistently aggregated across all types, reducing cognitive load on users.

Primer combines specialized ML models (embedding models, fine-tuned LMs, deep learning classifiers), and external LLM integrations, each optimized for its specific task rather than forcing everything through a single, general-purpose LLM. This modular architecture delivers particular advantages:

• computed enrichments persist across queries (no redundant processing)
• models swap independently without vendor lock-in
• specialized models provide measurable outputs with accessible weights
• structured extractions enable corpus-wide analytics like entity resolution across millions of documents

The result: superior accuracy and scale at a fraction of the computational cost of a monolithic LLM, with the control and flexibility that mission-critical systems demand.

‍