letsearch. A collection. An index. Relevant records.
letsearch provides retrieval. It embeds a text collection into an index and serves it as a search service, so a query finds records by meaning, not by shared words.
- Reads the question: A question in everyday words. Twelve records, none in focus yet.
- Finds by meaning: The closest record sharpens. It shares no word with the question.
- Refocuses: A new question moves the focus. Sharpness shows the score.
letsearch finds records by meaning, not by matching words. Twelve records sit out of focus; when a query is typed the closest record in meaning comes into focus. “The product runs very noisily” → “If the fan gets louder, lubricate the bearings.” (no shared words) “I want my money back, how long does it take?” → “Refund requests are settled within fourteen days.” (no shared words) Sharpness shows the similarity score computed with the ALTAI BGE-M3 distill model.
Connect applications to relevant text.
letsearch provides a semantic search layer by indexing text collections as vectors. It imports JSONL, Parquet, or Hugging Face datasets, embeds selected fields, and exposes indexed collections through a search service. The repository documents current API behavior and usage details.
Import text collections
Read JSONL and Parquet files, including datasets addressed through Hugging Face Hub paths.
Build a vector index
Select collection fields and an embedding model to create a searchable vector index.
Expose search to an application
Run a search service over an indexed collection and query it with the companion client.
From input to output.
- Input: A JSONL or Parquet collection with a text field and an embedding model suited to the task.
- Process: Import the collection, embed the selected field, build an index, and start the search service.
- Output: Indexed records matching a semantic query, available for an application to use.
In client projects.
ALTAI uses letsearch to build retrieval layers that connect document collections to applications. We prepare the source text and evaluation set, measure retrieval quality, and integrate results into the target workflow.
Run the first example.
Commands from the README. The binary was not run here.
# Download a compatible binary from the repository releases
./letsearch --help
# Example from the repository README
./letsearch index --collection-name test1 \
--index-columns context \
'hf://datasets/neural-bridge/rag-dataset-1200/**/*.parquet'
./letsearch serve -c test1
Spec sheet
- Category: Semantic search and retrieval
- Runtime: Rust / ONNX
- Interface: CLI / service
- License: Apache-2.0
- Source: GitHub, Documentation
- Models: Turkish embedding models
- Scope: Early stage. PDF import, MCP and incremental indexing are on the roadmap, not current features.