Local document search, end to end.

Problem

Search private working documents without sending them to a hosted index.

Greenlight is designed for developers and researchers who want a local-first retrieval stack they can inspect, test, and run through more than one interface.

Pipeline

One ingestion path, three ways to use it.

Implementation

Qwen embeddings with Qdrant retrieval.

• Ingests PDF, DOCX, TXT, and Markdown files.
• Chunks extracted text before embedding with a Qwen 3 SentenceTransformer model.
• Stores and retrieves vectors through Qdrant.
• Exposes workflows through a Rich-powered CLI, FastAPI endpoints, and Vue 3 UI.
• Ships with tests, documentation, automated setup, and Docker Compose orchestration.

Engineering decisions

Keep the retrieval core interface-agnostic.

The CLI, API, and browser interface sit on the same core pipeline. That makes it easier to test ingestion and retrieval without coupling correctness to one presentation layer.

Latency methodology

A “sub-100 ms” number needs a boundary.

Development runs have shown sub-100 ms semantic-search performance, but a responsible public benchmark must state corpus size, chunk count, embedding/index configuration, hardware, warm-up behavior, measurement boundary, and number of runs. Until that table is published, the number is treated as an engineering target rather than a universal claim.

Demonstration

Repository available; UI walkthrough next.

Limitations + next steps

Retrieval quality needs an explicit evaluation set.

Latency alone does not establish usefulness. Next steps include a labeled query set, retrieval metrics, corpus-scale experiments, chunking comparisons, and a documented hardware/runtime environment.

Lesson

Operational completeness makes a prototype credible.

Tests, setup, documentation, containers, and multiple interfaces turn an embedding experiment into a system someone else can actually evaluate.