Download serving/src/vector_index.rs from geraldaton20/alter_ego: direct link, hf CLI and curl.
- Browser
- Download file 2.76 kB
-
https://huggingface.co/geraldaton20/alter_ego/resolve/main/serving/src/vector_index.rs
- Command line
-
hf download hf://geraldaton20/alter_ego/serving/src/vector_index.rs
-
curl -L -o vector_index.rs https://huggingface.co/geraldaton20/alter_ego/resolve/main/serving/src/vector_index.rs
2.76 kB
| //! Ingestion-side vector index (plan E2.7 v1): exact cosine top-k over leaf | |
| //! vectors written during digestion. Ingestion-ONLY by construction - the | |
| //! recall path (recall.rs / BM25) never touches this module, honoring the | |
| //! PRD rule that the embedder's representations are not used at query time. | |
| //! Backend swap to true HNSW later keeps this API. | |
| use crate::error::{Error, Result}; | |
| pub struct VectorIndex { | |
| dim: usize, | |
| entries: Vec<(String, Vec<f32>)>, | |
| } | |
| impl VectorIndex { | |
| pub fn new(dim: usize) -> Self { | |
| Self { dim, entries: Vec::new() } | |
| } | |
| pub fn len(&self) -> usize { | |
| self.entries.len() | |
| } | |
| pub fn is_empty(&self) -> bool { | |
| self.entries.is_empty() | |
| } | |
| /// Approximate resident footprint (PRD section 7 RAM accounting). | |
| pub fn approx_bytes(&self) -> usize { | |
| self.entries.iter().map(|(id, v)| id.len() + v.len() * 4 + 48).sum() | |
| } | |
| pub fn add(&mut self, id: &str, vec: Vec<f32>) -> Result<()> { | |
| if self.dim == 0 { | |
| self.dim = vec.len(); | |
| } | |
| if vec.len() != self.dim { | |
| return Err(Error::Model(format!( | |
| "vector dim {} != index dim {}", | |
| vec.len(), | |
| self.dim | |
| ))); | |
| } | |
| self.entries.push((id.to_string(), vec)); | |
| Ok(()) | |
| } | |
| pub fn search(&self, query: &[f32], k: usize) -> Vec<(String, f32)> { | |
| if query.len() != self.dim || self.entries.is_empty() { | |
| return Vec::new(); | |
| } | |
| let qn: f32 = query.iter().map(|x| x * x).sum::<f32>().sqrt().max(1e-9); | |
| let mut scored: Vec<(String, f32)> = self | |
| .entries | |
| .iter() | |
| .map(|(id, v)| { | |
| let dot: f32 = query.iter().zip(v).map(|(a, b)| a * b).sum(); | |
| let vn: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt().max(1e-9); | |
| (id.clone(), dot / (qn * vn)) | |
| }) | |
| .collect(); | |
| scored.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal)); | |
| scored.truncate(k); | |
| scored | |
| } | |
| } | |
| mod tests { | |
| use super::*; | |
| fn nearest_neighbor_ranking() { | |
| let mut ix = VectorIndex::new(3); | |
| ix.add("cat", vec![1.0, 0.0, 0.0]).unwrap(); | |
| ix.add("dog", vec![0.9, 0.1, 0.0]).unwrap(); | |
| ix.add("car", vec![0.0, 0.0, 1.0]).unwrap(); | |
| let hits = ix.search(&[0.95, 0.05, 0.0], 2); | |
| assert_eq!(hits.len(), 2); | |
| assert_eq!(hits[0].0, "cat"); | |
| assert!(matches!(hits[1].0.as_str(), "dog")); | |
| assert!(ix.search(&[1.0, 0.0, 0.0], 3)[2].0 == "car"); | |
| assert!(ix.approx_bytes() > 0); | |
| assert!(ix.add("bad", vec![1.0]).is_err()); | |
| } | |
| } | |