> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rafflesia.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Architecture

> Compile once into immutable releases; serve stateless byte-range reads over object storage.

The engine splits the work in two: a heavy, offline **release compiler** and a
small, stateless **query executor**. Compile a corpus once; serve cheap,
reproducible computations many times.

## Immutable releases

A corpus is compiled into an immutable, content-addressed release. Every object
in it is referenced by URI, version, and SHA-256, so a
release is a fixed set of bytes that can never change underneath a query. A
mutable alias (like `stable`) points at one release; publishing swaps the
pointer atomically. The object bucket — not a database — is the source of truth.

## Byte-range reads

Object storage (GCS, S3, or Azure Blob) is the authoritative index. To answer a
query the engine opens a small metadata footer, prunes down to the relevant
shards, and then range-reads only the posting and payload bytes it actually
needs. Cost scales with the query, not the corpus.

## The query path

Candidate generation and verification run in a purpose-built engine; it does not
shell out to BLAST, MMseqs2, or Foldseek. A query seeds tokens, fetches
the matching posting ranges, aggregates candidates, then verifies survivors with
an exact **Smith-Waterman** alignment (`BLOSUM62`, affine gaps). The result
carries typed measurements plus the release, snapshot, and engine versions that
produced it.