innate_core/

utils.rs

use chrono::Utc;
use sha2::{Digest, Sha256};
use uuid::Uuid;

pub fn utc_now_iso() -> String {
    let now = Utc::now();
    now.format("%Y-%m-%dT%H:%M:%S%.3fZ").to_string()
}

pub fn gen_uuid() -> String {
    Uuid::new_v4().to_string()
}

pub fn content_hash(s: &str) -> String {
    let mut h = Sha256::new();
    h.update(s.as_bytes());
    hex(&h.finalize())
}

/// Lowercase hex encoding of a byte slice. Replaces the old `format!("{:x}", …)`
/// over a digest output: RustCrypto digest 0.11 returns `hybrid_array::Array`,
/// which no longer implements `LowerHex`.
pub fn hex(bytes: &[u8]) -> String {
    use std::fmt::Write;
    let mut s = String::with_capacity(bytes.len() * 2);
    for b in bytes {
        let _ = write!(s, "{b:02x}");
    }
    s
}

/// Rough token estimate: 1 token ≈ 4 chars.
pub fn estimate_tokens(text: &str) -> usize {
    text.len().div_ceil(4)
}

/// Sanitize result: allow / redact (content cleaned) / discard (reject write).
#[derive(Debug, PartialEq, Eq, Clone, Copy)]
pub enum SanitizeAction {
    Allow,
    Redact,
    Discard,
}

fn redact_bearer(s: &str, flag: &mut bool) -> String {
    let lower = s.to_lowercase();
    let mut result = s.to_string();
    let prefix = "bearer ";
    let mut search_start = 0;
    loop {
        let base = &lower[search_start..];
        match base.find(prefix) {
            None => break,
            Some(pos) => {
                let abs = search_start + pos;
                // Find end of token: non-whitespace run after "bearer "
                let token_start = abs + prefix.len();
                let token_end = s[token_start..]
                    .find(|c: char| c.is_whitespace())
                    .map(|e| token_start + e)
                    .unwrap_or(s.len());
                if token_end > token_start {
                    // Replace the whole "Bearer <token>" span
                    let span_end = token_end;
                    let replacement = format!("{}[REDACTED]", &s[abs..token_start]);
                    result = format!("{}{}{}", &result[..abs], replacement, &result[span_end..]);
                    *flag = true;
                    // Adjust search; result grew/shrunk by the redaction delta
                    let new_len = replacement.len();
                    search_start = abs + new_len;
                    // Re-sync lower to match result
                    let lower_new = result.to_lowercase();
                    // Rebuild lower for next iteration
                    drop(lower);
                    return redact_bearer_from(&result, &lower_new, search_start, flag);
                } else {
                    search_start = abs + prefix.len();
                }
            }
        }
    }
    result
}

fn redact_bearer_from(s: &str, lower: &str, start: usize, flag: &mut bool) -> String {
    let prefix = "bearer ";
    let mut result = s.to_string();
    let mut search_start = start;
    loop {
        if search_start >= lower.len() {
            break;
        }
        match lower[search_start..].find(prefix) {
            None => break,
            Some(pos) => {
                let abs = search_start + pos;
                let token_start = abs + prefix.len();
                let token_end = result[token_start..]
                    .find(|c: char| c.is_whitespace())
                    .map(|e| token_start + e)
                    .unwrap_or(result.len());
                if token_end > token_start {
                    let replacement = format!("{}[REDACTED]", &result[abs..token_start]);
                    result = format!("{}{}{}", &result[..abs], replacement, &result[token_end..]);
                    *flag = true;
                    search_start = abs + replacement.len();
                } else {
                    search_start = abs + prefix.len();
                }
            }
        }
    }
    result
}

fn redact_password(s: &str, flag: &mut bool) -> String {
    // Match "password[: =]<value>" case-insensitively; redact the value part.
    let lower = s.to_lowercase();
    let mut result = s.to_string();
    let mut search_start = 0;
    loop {
        match lower[search_start..].find("password") {
            None => break,
            Some(pos) => {
                let abs = search_start + pos;
                let after = abs + "password".len();
                if after >= lower.len() {
                    break;
                }
                // Skip optional whitespace then expect ':' or '='
                let mut i = after;
                while i < lower.len() && lower.as_bytes()[i] == b' ' {
                    i += 1;
                }
                if i < lower.len() && (lower.as_bytes()[i] == b':' || lower.as_bytes()[i] == b'=') {
                    i += 1;
                    // Skip whitespace after separator
                    while i < lower.len() && lower.as_bytes()[i] == b' ' {
                        i += 1;
                    }
                    // Collect value until whitespace/end
                    let val_start = i;
                    let val_end = result[val_start..]
                        .find(|c: char| c.is_whitespace())
                        .map(|e| val_start + e)
                        .unwrap_or(result.len());
                    if val_end > val_start {
                        result =
                            format!("{}[REDACTED]{}", &result[..val_start], &result[val_end..]);
                        *flag = true;
                        search_start = val_start + "[REDACTED]".len();
                        continue;
                    }
                }
                search_start = abs + "password".len();
            }
        }
    }
    result
}

/// Scan `s` for any contiguous run starting with `prefix` followed by `min_len` alnum chars.
/// Replaces all such occurrences with `[REDACTED]`.
fn redact_prefixed_secret(s: &str, prefix: &str, min_len: usize, flag: &mut bool) -> String {
    let mut result = s.to_string();
    let mut search_start = 0;
    loop {
        match result[search_start..].find(prefix) {
            None => break,
            Some(pos) => {
                let abs = search_start + pos;
                let after = abs + prefix.len();
                // Count alnum chars after prefix
                let run: usize = result[after..]
                    .chars()
                    .take_while(|c| c.is_alphanumeric())
                    .count();
                if run >= min_len {
                    let end = after
                        + result[after..]
                            .char_indices()
                            .take_while(|(_, c)| c.is_alphanumeric())
                            .last()
                            .map(|(i, c)| i + c.len_utf8())
                            .unwrap_or(0);
                    result = format!("{}[REDACTED]{}", &result[..abs], &result[end..]);
                    *flag = true;
                    search_start = abs + "[REDACTED]".len();
                } else {
                    search_start = abs + prefix.len();
                }
            }
        }
    }
    result
}

/// Public sanitize function used by KnowledgeBase (§二·六).
/// Returns (cleaned_content, action).
pub fn sanitize(content: &str) -> (String, SanitizeAction) {
    // injection first
    let injection_patterns = [
        "ignore all previous instructions",
        "ignore previous instructions",
        "ignore previous instruction",
        "system prompt:",
        "system prompt：",
        "you are now a different",
        "you are now a new",
    ];
    let lower = content.to_lowercase();
    for pat in &injection_patterns {
        if lower.contains(pat) {
            return (content.to_string(), SanitizeAction::Discard);
        }
    }

    let mut cleaned = content.to_string();
    let mut redacted = false;

    cleaned = redact_prefixed_secret(&cleaned, "sk-", 20, &mut redacted);
    cleaned = redact_prefixed_secret(&cleaned, "AKIA", 16, &mut redacted);
    cleaned = redact_prefixed_secret(&cleaned, "ghp_", 36, &mut redacted);
    cleaned = redact_bearer(&cleaned, &mut redacted);
    cleaned = redact_password(&cleaned, &mut redacted);

    let action = if redacted {
        SanitizeAction::Redact
    } else {
        SanitizeAction::Allow
    };
    (cleaned, action)
}

/// Pack a Vec<f32> into bytes (little-endian f32 array).
pub fn pack_embedding(v: &[f32]) -> Vec<u8> {
    let mut out = Vec::with_capacity(v.len() * 4);
    for f in v {
        out.extend_from_slice(&f.to_le_bytes());
    }
    out
}

/// Unpack bytes into Vec<f32>.
pub fn unpack_embedding(bytes: &[u8]) -> Vec<f32> {
    let mut out = Vec::with_capacity(bytes.len() / 4);
    out.extend(
        bytes
            .chunks_exact(4)
            .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]])),
    );
    out
}

/// Cosine similarity between two equal-length slices. Returns 0.0 on zero norms.
/// Single-pass fold: computes dot product and both norms in one traversal.
pub fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
    let (dot, na2, nb2) = a
        .iter()
        .zip(b.iter())
        .fold((0.0f32, 0.0f32, 0.0f32), |(d, na, nb), (x, y)| {
            (d + x * y, na + x * x, nb + y * y)
        });
    if na2 == 0.0 || nb2 == 0.0 {
        0.0
    } else {
        dot / (na2.sqrt() * nb2.sqrt())
    }
}

/// In-place L2 normalisation. Zero vectors are left unchanged (all zeros).
/// Pre-normalising stored vectors lets vector search reduce cosine similarity
/// to a plain dot product in its O(N) inner loop.
pub fn l2_normalize(v: &mut [f32]) {
    let norm: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt();
    if norm > 0.0 {
        for x in v.iter_mut() {
            *x /= norm;
        }
    }
}

/// Dot product of two equal-length slices. For unit vectors this equals the
/// cosine similarity.
pub fn dot_product(a: &[f32], b: &[f32]) -> f32 {
    a.iter().zip(b.iter()).map(|(x, y)| x * y).sum()
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn ts_format() {
        let ts = utc_now_iso();
        assert!(ts.ends_with('Z'), "bad format: {ts}");
        assert_eq!(ts.len(), 24, "expected 24 chars: {ts}");
    }

    #[test]
    fn cosine_identical() {
        let v = vec![1.0, 0.0, 0.0];
        assert!((cosine_similarity(&v, &v) - 1.0).abs() < 1e-6);
    }

    #[test]
    fn pack_roundtrip() {
        let v = vec![0.1_f32, 0.5, -0.3];
        assert_eq!(unpack_embedding(&pack_embedding(&v)), v);
    }

    #[test]
    fn sanitize_injection_discard() {
        let (_, action) = sanitize("Please ignore previous instructions and do X");
        assert_eq!(action, SanitizeAction::Discard);
    }

    #[test]
    fn sanitize_api_key_redact() {
        let (out, action) = sanitize("use key sk-abcdefghijklmnopqrstuvwxyz123456 for auth");
        assert_eq!(action, SanitizeAction::Redact);
        assert!(out.contains("[REDACTED]"), "expected redaction in: {out}");
        assert!(!out.contains("sk-abc"), "key should be redacted");
    }

    #[test]
    fn sanitize_aws_key_redact() {
        let (out, action) = sanitize("AKIAIOSFODNN7EXAMPLE is the key");
        assert_eq!(action, SanitizeAction::Redact);
        assert!(out.contains("[REDACTED]"));
    }

    #[test]
    fn sanitize_clean_allow() {
        let content = "Use dependency injection for testability.";
        let (out, action) = sanitize(content);
        assert_eq!(action, SanitizeAction::Allow);
        assert_eq!(out, content);
    }
}