hypersteeldb 0.3.2

//! Export **real** typed spans and their embeddings, so the browser demo can run the actual ontology
//! discovery instead of a substitute for it.
//!
//!   `cargo run --release --features onnx,embed --bin export_spans -- <corpus_dir> <out.json> [max_docs]`
//!
//! Two stages of the pipeline are inherently model-bound and cannot compile to WebAssembly: the SPO tagger
//! (ONNX Runtime) and the model2vec embedder (its tokenizer links a C regex library). Everything *after*
//! them — per-kind k-means, Sinkhorn optimal transport, the MECE gate — is pure arithmetic and runs anywhere.
//!
//! So those two stages run here, natively, and their output is shipped as data. The browser then executes the
//! genuine discovery algorithm on genuine typed spans, rather than approximating the whole thing with a
//! cheaper geometry. The model is an input; the algorithm is what the demo is about.
//!
//! Following the reference `spo_sinkhorn.py`: `QTY` is routed to numeric rules and `IGNORE` to boilerplate,
//! so neither enters the semantic codebook. Only `ENT`, `REL`, `GEO` and `TIME` are clustered.

use std::collections::BTreeMap;
use std::path::PathBuf;

/// The kinds that form the facet axes. Everything else is routed away rather than clustered.
const CODEBOOK_KINDS: &[&str] = &["ENT", "REL", "GEO", "TIME"];

fn sentences(text: &str) -> Vec<String> {
    text.split(['.', '\n', ';', '!', '?'])
        .map(|s| s.trim())
        .filter(|s| s.split_whitespace().count() >= 3)
        .map(|s| s.to_string())
        .collect()
}

fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    let args: Vec<String> = std::env::args().collect();
    let dir = args.get(1).cloned().unwrap_or_else(|| "benchmark_corpus".into());
    let out_path = args.get(2).cloned().unwrap_or_else(|| "spans.json".into());
    let max_docs: usize = args.get(3).and_then(|s| s.parse().ok()).unwrap_or(24);

    let ml = steeldb::paths::model_dir("step0_bundle_ml", "STEELDB_ML_BUNDLE", "spo.onnx")
        .ok_or("SPO tagger bundle not found (models/step0_bundle_ml/spo.onnx)")?;
    let m2v_dir = steeldb::paths::model_dir("model2vec", "STEELDB_MODEL2VEC", "potion.f32")
        .ok_or("model2vec not found (models/model2vec/potion.f32)")?;

    let mut tagger = steeldb::text::tagger::SpoTagger::load(&ml)?;
    let embedder = steeldb::text::Model2Vec::load(&m2v_dir)?;
    eprintln!("tagger: {} · embedder: {} dims", ml.display(), embedder.dim());

    // read the corpus, splitting on the same `---` separator the demo uses
    let path = PathBuf::from(&dir);
    let raw = if path.is_file() {
        std::fs::read_to_string(&path)?
    } else {
        let mut files: Vec<PathBuf> = std::fs::read_dir(&path)?
            .filter_map(|e| e.ok().map(|e| e.path()))
            .filter(|p| p.extension().map(|x| x == "md" || x == "txt").unwrap_or(false))
            .collect();
        files.sort();
        files
            .iter()
            .take(max_docs)
            .filter_map(|p| std::fs::read_to_string(p).ok())
            .collect::<Vec<_>>()
            .join("\n---\n")
    };
    let docs: Vec<String> = raw
        .split("\n---")
        .map(|d| d.trim().to_string())
        .filter(|d| !d.is_empty())
        .take(max_docs)
        .collect();
    eprintln!("documents: {}", docs.len());

    // tag every sentence, keep the first occurrence of each surface per kind
    let mut seen: BTreeMap<(String, String), usize> = BTreeMap::new();
    let mut routed_away: BTreeMap<String, usize> = BTreeMap::new();
    let mut records: Vec<serde_json::Value> = Vec::new();
    let mut numeric_rules: Vec<serde_json::Value> = Vec::new();

    for (di, doc) in docs.iter().enumerate() {
        for sentence in sentences(doc) {
            let spans = match tagger.tag(&sentence) {
                Ok(s) => s,
                Err(e) => {
                    eprintln!("  tag failed: {e}");
                    continue;
                }
            };
            for sp in spans {
                // recover the whole words the prediction touched, rather than the wordpiece slice
                let (ss, se) = steeldb::spans::snap_to_words(&sentence, sp.start, sp.end);
                let surface = sentence
                    .get(ss..se)
                    .map(|t| t.trim().to_string())
                    .unwrap_or_else(|| sp.text.trim().to_string());
                // a fragment that snapping could not resolve is not worth clustering
                if surface.chars().count() < 3 || !surface.chars().any(|c| c.is_alphanumeric()) {
                    continue;
                }
                if !CODEBOOK_KINDS.contains(&sp.kind.as_str()) {
                    *routed_away.entry(sp.kind.clone()).or_default() += 1;
                    // A quantity does not vanish when it leaves the codebook — it becomes a numeric field with
                    // a value, which is what makes range predicates possible. Export where it went, so the
                    // claim "routed to the numeric rules" can be shown rather than asserted.
                    if sp.kind == "QTY" {
                        for (qs, qe, field) in steeldb::emergent::quantity_spans(&surface) {
                            let text = &surface[qs..qe];
                            // the leading sign is part of the value
                            let digits: String = text
                                .chars()
                                .enumerate()
                                .take_while(|(i, c)| c.is_ascii_digit() || *c == '.' || (*i == 0 && *c == '-'))
                                .map(|(_, c)| c)
                                .collect();
                            if let Ok(v) = digits.parse::<f64>() {
                                numeric_rules.push(serde_json::json!({
                                    "span": surface, "field": field, "value": v, "doc": di,
                                }));
                            }
                        }
                    }
                    continue;
                }
                let key = (sp.kind.clone(), surface.to_lowercase());
                let count = seen.entry(key.clone()).or_insert(0);
                *count += 1;
                if *count > 1 {
                    continue; // one record per distinct surface+kind; the count carries the frequency
                }
                let Some(vec) = embedder.embed(&surface) else { continue };
                // round to 4 decimals: the plan is unaffected and the payload roughly halves
                let v: Vec<f32> = vec.iter().map(|x| (x * 10_000.0).round() / 10_000.0).collect();
                records.push(serde_json::json!({
                    "text": surface,
                    "kind": sp.kind,
                    "doc": di,
                    "vec": v,
                }));
            }
        }
    }

    // attach the observed frequency, so the browser can weight or filter by support
    for r in records.iter_mut() {
        let k = (
            r["kind"].as_str().unwrap_or("").to_string(),
            r["text"].as_str().unwrap_or("").to_lowercase(),
        );
        r["count"] = serde_json::json!(seen.get(&k).copied().unwrap_or(1));
    }

    let mut by_kind: BTreeMap<&str, usize> = BTreeMap::new();
    for r in &records {
        *by_kind.entry(r["kind"].as_str().unwrap_or("?")).or_default() += 1;
    }

    let payload = serde_json::json!({
        "documents": docs.len(),
        "dim": embedder.dim(),
        "codebook_kinds": CODEBOOK_KINDS,
        "spans": records,
        "routed_away": routed_away,
        "numeric_rules": numeric_rules,
        "source": "SPO tagger (spo.onnx) + model2vec embeddings, exported natively",
    });
    std::fs::write(&out_path, serde_json::to_string(&payload)?)?;

    eprintln!("spans kept for the codebook:");
    for (k, n) in &by_kind {
        eprintln!("  {k:<6} {n}");
    }
    eprintln!("routed away (never enter the codebook):");
    for (k, n) in &routed_away {
        eprintln!("  {k:<6} {n}");
    }
    let bytes = std::fs::metadata(&out_path).map(|m| m.len()).unwrap_or(0);
    eprintln!("wrote {} ({} KB)", out_path, bytes / 1024);
    Ok(())
}