mirror of
https://github.com/hanzoai/semantic.git
synced 2026-10-30 15:04:19 +00:00
zeekayandHanzo Dev
1cd0cd7644
The first commit was the shape. This is the implementation: every stage of the pipeline, the graph and its measures, the type system over it, entailment, deduplication and conflict resolution, the decision layer, the stores and the export formats. Twenty-six thousand lines of Go and sixteen thousand of tests, against no dependency but the standard library. Where the Python reaches for torch, spacy or a database driver, the Go states an interface and reports what is missing rather than pretending. A caller can tell "this build cannot serve that" from "nobody has heard of it", which is the difference between a port and a stub. The deterministic paths are complete and are what the tests exercise; the model-shaped ones are one adapter away. Porting a bug is not parity, so several places now disagree with the original on purpose. The JSON repair that closed a truncated reply counted each bracket kind separately and closed them backwards on the one input it exists for; it now works off a stack and ignores brackets inside strings. The entity patterns joined words with \s, so a heading ran into the sentence beneath it and became one name while the real relation was lost; a name is separated by horizontal space now. Eigenvector centrality ran plain power iteration, which on a bipartite graph swings between the two ends of a sign-paired eigenvalue forever and reported every node in a star as equally central; it iterates with A+I, which moves no eigenvector and settles. Similarity summed its weighted parts in map order, and float addition is not associative, so the same comparison scored differently between runs and duplicate detection returned a different answer each time. An inferred ontology named object-property endpoints by raw source type and data-property endpoints by class, so the schema referred to classes it did not contain. A property with no IRI of its own was minted as a class and came back Pascal-cased, renaming the term on the way out of the document. The command built a pipeline with every stage nil and panicked on each invocation. It is now the composition the package documents, and Run reports a missing ingester instead of dereferencing it. Co-authored-by: Hanzo Dev <[email protected]>
87 lines
2.6 KiB
Go
87 lines
2.6 KiB
Go
package split
|
|
|
|
import (
|
|
"context"
|
|
"errors"
|
|
"fmt"
|
|
"math"
|
|
"strings"
|
|
"unicode/utf8"
|
|
|
|
"github.com/hanzoai/semantic"
|
|
)
|
|
|
|
// Embed turns texts into vectors. Semantic calls it once per document with
|
|
// every sentence in it, so an implementation is free to batch, and the context
|
|
// is the one the split was started with, so a slow model is cancellable.
|
|
type Embed func(ctx context.Context, texts []string) ([][]float32, error)
|
|
|
|
// Semantic keeps consecutive sentences together while they mean similar
|
|
// things and cuts where they stop. What counts as similar is the cosine of the
|
|
// angle between adjacent sentence vectors; Threshold is where the cut falls.
|
|
// The vectors come from Embed, which this package never supplies: no splitter
|
|
// here reaches the network on its own.
|
|
type Semantic struct {
|
|
Embed Embed // required
|
|
Threshold float64 // cut when adjacent similarity falls below this
|
|
Size int // runes per chunk; zero or less lets similarity alone decide
|
|
}
|
|
|
|
// Split cuts d where the meaning of its sentences shifts. It returns an error
|
|
// if Embed is missing, fails, or answers with the wrong number of vectors.
|
|
func (s Semantic) Split(ctx context.Context, d semantic.Doc) ([]semantic.Chunk, error) {
|
|
if err := ctx.Err(); err != nil {
|
|
return nil, err
|
|
}
|
|
if d.Text == "" {
|
|
return nil, nil
|
|
}
|
|
if s.Embed == nil {
|
|
return nil, errors.New("split: Semantic needs an Embed")
|
|
}
|
|
units := sentences(d.Text)
|
|
if len(units) < 2 {
|
|
return chunks(d, units), nil
|
|
}
|
|
texts := make([]string, len(units))
|
|
for i, u := range units {
|
|
texts[i] = strings.TrimSpace(d.Text[u.start:u.end])
|
|
}
|
|
vectors, err := s.Embed(ctx, texts)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("split: embedding %d sentences: %w", len(texts), err)
|
|
}
|
|
if len(vectors) != len(units) {
|
|
return nil, fmt.Errorf("split: got %d vectors for %d sentences", len(vectors), len(units))
|
|
}
|
|
size := s.Size
|
|
if size <= 0 {
|
|
size = utf8.RuneCountInString(d.Text) + 1
|
|
}
|
|
p := pack{
|
|
size: size,
|
|
cut: func(i int) bool { return cosine(vectors[i-1], vectors[i]) < s.Threshold },
|
|
}
|
|
return chunks(d, p.spans(d.Text, units)), nil
|
|
}
|
|
|
|
// cosine measures the angle between two vectors, on a scale where 1 is the
|
|
// same direction and 0 is unrelated. Vectors of different width, and vectors
|
|
// with no length, have no angle between them, and count as unrelated.
|
|
func cosine(a, b []float32) float64 {
|
|
if len(a) != len(b) || len(a) == 0 {
|
|
return 0
|
|
}
|
|
var dot, na, nb float64
|
|
for i := range a {
|
|
x, y := float64(a[i]), float64(b[i])
|
|
dot += x * y
|
|
na += x * x
|
|
nb += y * y
|
|
}
|
|
if na == 0 || nb == 0 {
|
|
return 0
|
|
}
|
|
return dot / (math.Sqrt(na) * math.Sqrt(nb))
|
|
}
|