Expand description
§QQL Edge — zero-network vector search
Combines fastembed-rs for local ONNX embedding inference with qdrant-edge for in-process HNSW vector search. No network hops, no external services, no API keys — unless you choose an external provider.
§Embedder options
| Function | Embedder | Network? |
|---|---|---|
local_executor | fastembed (ONNX, local CPU) | ❌ none |
local_executor_with_options | fastembed + model/cache selection | ❌ none |
[http_executor] | OpenAI-compatible HTTP endpoint | ✅ provider only |
custom_executor | Any Arc<dyn Embedder> | up to you |
§Quick start
use qql_edge::local_executor;
use qql::executor::OnError;
let mut executor = local_executor("/tmp/qql-edge-data", false)?;
let resp = executor.execute("CREATE COLLECTION docs HYBRID", OnError::Stop).await?;§JSON scope (“envelope removal”)
Responses are fully typed (ExecData, no JSON envelope). serde_json::Value
remains only where the engine API requires it: point payloads
(PointStruct::new), typed-spec introspection maps, the turbo
quantization builder map, and formula DEFAULTS (edge formulas are JSON).
Value-based parity helpers are test-only (#[cfg(test)]).
Structs§
- Edge
Qdrant - In-process backend: one
qdrant_edgeshard per collection underbase_path, zero network. Batch methods fan out to individual routes (no native batch RPC). - Embedding
Model Info - Public description of a local ONNX embedding model.
- Fast
Embedder - Local fastembed embedder: dense ONNX model plus optional sparse, multi (ColBERT), image, and cross-encoder slots, with wire-compatible BM25 fallback.
- Fast
Embedder Options - Options for constructing a
FastEmbedder. - Local
Executor Options - Options for
local_executor_with_options. - Shard
Summary - Counts collected from a snapshot-seeded shard.
Functions§
- custom_
executor - Build an edge
Executorwith a fully customEmbedder. - custom_
executor_ with_ dimension - Build an edge
Executorwith a custom embedder and an optional explicit dense dimension. The embedder-reported dimension takes precedence. - custom_
executor_ with_ storage - Like
custom_executor_with_dimension, with an explicit WAL segment capacity (bytes) applied to every shard the executor opens or creates;Nonekeeps the engine default. - inspect_
shard - Load the shard at
shard_dir, read its counts, and close it again. - list_
embedding_ models - List dense text models, offline sparse models (SPLADE / BGE-M3), multi (BGE-M3 / ColBERT), and image models (CLIP vision, …) that fastembed can load.
- local_
executor - Build a fully-local
Executorbacked by fastembed-rs and qdrant-edge. - local_
executor_ with_ options - Build a fully-local
Executorwith explicit model / cache options. - resolve_
embedding_ model - Resolve a user-facing dense model string to an
EmbeddingModel. - resolve_
image_ model - Resolve offline image model (CLIP vision, etc.).
- resolve_
multi_ model - Resolve offline multi model (
bge-m3, HF code, etc.). - unpack_
snapshot - Unpack a downloaded shard-snapshot archive into
target. - wal_
segment_ capacity_ bytes - Convert a MiB WAL segment capacity into the byte count
LocalExecutorOptions::wal_segment_capacityexpects.