BiVelio Embedding Layer Algorithm
The best evidence, with fewer embeddings and fewer tokens.
BV-EMLA turns documents and a query into a minimal package of relevant evidence. It reduces the cost of creating embeddings, storing indexes and sending context to the model, while preserving the identity, versions, metadata and traceability of each document. It does not necessarily generate an answer: it prepares the best evidence so that any application —with OpenAI, Anthropic, Gemini or a local model— can use it.
BV-EMLA is designed as a fully standalone product. It does not need BiVelio's Brain, nor an account, nor its databases or its internal models. BiVelio will later be one of its consumers, not the environment it needs to run.
- 1
Ingestion: it normalizes, deduplicates in a verified way and chunks the documents.
- 2
Representation: it reuses embeddings from cache or creates only the missing ones.
- 3
Adaptive retrieval: it chooses the cheapest plan that reaches sufficient evidence.
- 4
Composition: it delivers the minimal context that covers the query, with traces and metrics.
EMLA Compile
It transforms a collection of documents into an efficient and verifiable representation: it normalizes, detects duplicates and near-duplicates in a verified way, chunks by structure and reuses or creates embeddings. Ten copies of the same manual share a single technical representation, without losing the identity of each document.
EMLA Resolve
It receives a query and chooses how much effort it needs. A simple question uses a cheap search; a complex one adds dense retrieval, reranking and expansion. You pay the high cost only when the query demands it, and the result is the smallest set that covers the whole question.
Fewer repeated embeddings and less storage.
Fewer context tokens sent to the model.
Interchangeable indexes and providers; local mode with no external infrastructure.
Traceability and metrics for every retrieval.
In design as a standalone library (Python SDK and HTTP API). It will first be validated with neutral benchmarks; then it will be integrated into BiVelio through an adapter.
Algorithm in design. Performance and savings figures will be published once we complete the benchmarks; until then we make no claims about percentages.