bankai - v0.3.0

bankai/vector_bridge

Vector retrieval over aarondb’s HNSW index.

Mnesia is Bankai’s source of truth. This module builds a short-lived index from Bankai documents for one command, keeping aarondb behind a Bankai-shaped result type and the embedding backend behind bankai/embed.

The index topology is deliberately deterministic. The managed projection is keyed by the committed Mnesia offset, so queries reuse a daemon-local HNSW graph until committed membership changes; direct search stays available for finite-corpus tests and one-shot callers.

Types

Document

</> 
pub type Document {
  Document(kind: String, id: String, text: String)
}

Constructors 

  • Document(kind: String, id: String, text: String)

Match

</> 
pub type Match {
  Match(kind: String, id: String, score: Float)
}

Constructors 

  • Match(kind: String, id: String, score: Float)

ProjectionStatus

</> 
pub type ProjectionStatus {
  ProjectionStatus(
    last_applied_offset: Int,
    document_count: Int,
    health: projection_index.Health,
    generation: Int,
    backend: String,
  )
}

Constructors 

  • ProjectionStatus(
  last_applied_offset: Int,
  document_count: Int,
  health: projection_index.Health,
  generation: Int,
  backend: String,
)

Values

backend

</> 
pub fn backend() -> String
</> 
pub fn exact_search(
  docs: List(Document),
  query: String,
  threshold: Float,
  limit: Int,
) -> List(Match)

Return the exact finite-corpus oracle for Bankai’s lexical vectors.

This exists for verification and benchmarks. The CLI intentionally keeps the HNSW path: exact search is exhaustive rather than an interactive retrieval strategy.

</> 
pub fn projected_exact_search(
  workspace: String,
  offset: Int,
  docs: List(Document),
  query: String,
  threshold: Float,
  limit: Int,
) -> Result(List(Match), String)

Exact oracle over the same managed projection corpus. It exists solely for verification and benchmark parity; user-facing commands use HNSW.

</> 
pub fn projected_search(
  workspace: String,
  offset: Int,
  docs: List(Document),
  query: String,
  threshold: Float,
  limit: Int,
) -> Result(List(Match), String)

Query the managed daemon-local HNSW projection at a committed source offset. Reusing an identical offset never rebuilds the graph; an offset advance produces a fresh deterministic generation before results are returned.

projection_status

</> 
pub fn projection_status(
  workspace: String,
) -> Result(ProjectionStatus, String)

reset_projection_for_test

</> 
pub fn reset_projection_for_test(
  workspace: String,
) -> Result(Nil, String)
</> 
pub fn search(
  docs: List(Document),
  query: String,
  threshold: Float,
  limit: Int,
) -> List(Match)

Build an in-memory HNSW index and return ranked document matches.

threshold is cosine similarity because embed.embed returns normalized vectors. Empty queries and non-positive limits intentionally return no matches rather than making the index treat a zero vector as relevant.

warm_projection

</> 
pub fn warm_projection(
  workspace: String,
  offset: Int,
  docs: List(Document),
) -> Result(Int, String)

Build the managed HNSW projection now so the cache table is owned by the calling (long-lived daemon) process. An ETS table dies with its owner; created lazily inside a per-connection handler, the projection would be discarded after every request.

✨ Search Document