TLDR-G
TLDR-G provides a knowledge rendering engine that builds a two-scale graph of atoms and passages from documents, with typed connections across structural, temporal, causal, and authorial axes. The engine enforces a hard token budget, preserving verbatim quotations where they matter and summarizing where they do not, achieving parity with full-context baselines at roughly 88% fewer tokens on legal multi-hop tasks. It runs fully on-device with single-file SQLite storage and is model-agnostic, supporting any user-supplied local, EU-hosted, or open-weights model.
- Legal Technology
- Artificial Intelligence
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Retrieval-augmented generation systems that rely on chunk-based retrieval often lose the contextual connections between pieces of information, undermining the truth they were meant to carry. They also tend to dump excessive context into prompts, inflating token costs and degrading answer precision.
Solution
TLDR-G provides a knowledge rendering engine that builds a two-scale graph of atoms and passages from documents, with typed connections across structural, temporal, causal, and authorial axes. For each question, the engine descends this graph and allocates a hard token budget across what is relevant, preserving verbatim quotations where they matter and summarizing where they do not. The engine runs fully on-device with single-file SQLite storage, so data never has to leave the user's machine. It is model-agnostic: the deterministic core needs no LLM at all, and users can bring their own local, EU-hosted, or open-weights model. Every rendered passage carries verifiable provenance back to its source, supported by a built-in verify step over signed attestations that third parties can confirm offline without requiring trust.
Target Audience
Legal professionals, researchers, and knowledge-intensive teams who need verifiable, token-efficient retrieval over large document corpora — particularly those with strict data sovereignty or confidence requirements that necessitate on-device processing and third-party auditability.
Features
- Two-scale graph representation of documents (atoms and passages) with typed connections across structural, temporal, causal, and authorial axes
- Hard token budget enforcement that protects high-signal context instead of truncating it, achieving parity with full-context baselines at roughly 88% fewer tokens on legal multi-hop tasks
- On-device, single-file (SQLite) storage with no data egress by default, described as "EU-resident by construction"
- Built-in verify step with signed attestations that lets third parties confirm cited text exists in the original source offline
- Model-agnostic design with a deterministic core that requires no LLM and supports any user-supplied model (local, EU-hosted, or open-weights)
- Open SDK under Apache-2.0 providing stable integration contracts and verification surface, while the engine itself is distributed as a free compiled binary