This page repeats the key product facts, benchmark metrics, limits, and collaboration context without relying on images.
View product pageTrajbl is a CPU-first, model-free, post-retrieval context compressor for RAG systems, AI agents, knowledge-base chat, vault chat, and private AI assistants.
Trajbl receives context that an existing retriever has already selected. It keeps complete evidence sentences, drops irrelevant surrounding text, and sends a smaller packet to the LLM.
Trajbl is not an LLM, not a vector database, and not a replacement for retrieval. Its intended production role is: existing retrieval top-k to Trajbl to LLM.
Raw BGE top50 remains the uncompressed accuracy ceiling in the current benchmark. Trajbl is evaluated as the compression layer over retrieved inputs.
On controlled public benchmark suites such as LongBench-style multi-hop QA and WixQA-style support workloads, Trajbl showed stronger quality-per-token than Microsoft LLMLingua2.
In the long-context signal, full-context RAG reached 0.587 F1 with 10,593 mean context tokens. Trajbl reached 0.400 F1 with 1,313 mean context tokens. LLMLingua2 reached 0.233 F1 at a similar token budget.
| Method | Quality | Quality retained | Mean context tokens | Token reduction |
|---|---|---|---|---|
| Full RAG reference | 0.587 F1 | 100% | 10,593 | 0% |
| Trajbl | 0.400 F1 | 68% | 1,313 | ~88% |
| LLMLingua2 | 0.233 F1 | 40% | 1,405 | ~87% |
Trajbl can reduce paid LLM input-token usage in applications that repeatedly send retrieved context to a model. Examples include Claw-like coding agents, Claude Code style tools, Codex-like tools, Obsidian vault chat, Open WebUI document chat, AnythingLLM, PrivateGPT, customer-support knowledge bases, and enterprise RAG systems.
A simple interpretation: if a system normally sends a large retrieved context packet to the LLM, Trajbl can trim that packet before generation while keeping the evidence sentences most likely to answer the question.
The current public-facing Trajbl benchmark evidence is not a universal final benchmark. The next validation steps are broader public benchmark runs, an infrastructure benchmark against LLMLingua2 on CPU-only hardware, and product pilots on real RAG and agent workflows.
The current QA and judge model was Gemini 2.5 Flash Lite through Vertex AI. Accuracy may change with stronger QA models, larger samples, different corpora, or different retrieval quality.
Trajbl is the current primary product focus. CurveMemory is presented as a separate promising research track for geometric routing and memory classification.
Trajbl is seeking pilots, strategic partners, investment, licensing discussions, and validation on real customer workloads.