Files
zeekay 25011a2085 brain: 100% fortemi parity β€” every algorithm shipped across all 4 runtimes
Closed every remaining 🟑 / πŸ”΅ row in PARITY.md. Each algorithm FortΓ©mi
ships now has a real implementation in @hanzo/bot-memory v2.0.0 with
matching ports in hanzo-memory (Python), bot-go (Go), and hanzoai/mcp
(Rust). Cross-runtime tests: 121 TS + 53 Python + 58 Go + 38 Rust = 270
passing.

Landed under packages/memory/

  Retrieval & ranking
    fusion       RRF (k=20 BEIR) + RSF (Weaviate v1.24) + adaptive k +
                 adaptive FTS/semantic weights
    rerank       MMR (Carbonell & Goldstein 1998), cosine
    dedup        chunk-aware deduplication
    script       Unicode script detection (Latin/CJK/emoji/Cyrillic/
                 Arabic/Hebrew/Greek/Devanagari)
    fts          feature-flag matrix, CJK bigrams, emoji trigrams,
                 websearch_to_tsquery parser, SQLite FTS5 MATCH renderer
    embed        embedding registry, asymmetric E5 prefixes, Matryoshka
                 truncation, coarse-dim picker
    two-stage    coarse β†’ fine retrieval via MRL
    federated    cross-archive search with RRF fusion
    multi-memory X-Hanzo-Memory routing (X-Fortemi-Memory alias)
    filters      composable tag/temporal/collection/scope predicates
    temporal     UUIDv7 floor/ceiling + named ranges
    spatial      Haversine + bbox + in-box (PostGIS fallback)

  Knowledge & graph
    graph        normalize β†’ SNN β†’ PFNET-∞ β†’ Louvain maintenance pipeline
    skos         W3C SKOS concept tree (broader/narrower closures,
                 Turtle export)
    fair         Dublin Core XML + JSON-LD
    doc-types    document type registry + auto-detect (filename/MIME/
                 content sniff)
    captions     WebVTT / SRT / RTTM renderers
    sprite       thumbnail sprite-sheet WebVTT cue builder
    code-ast     language-aware syntactic chunker
    email        RFC 2822 / MIME parser (multipart, qp, base64, HTML strip)
    exif         EXIF parser (camera / time / GPS)

  Operational
    events       EventBus + SSE renderer + webhook HMAC signer
    tokenizer    BPE-style estimator + truncate-to-budget
    eval         MRR / recall@k / precision@k / NDCG@k / benchmark
    hardware     tier autodetect + per-tier model recommender
    circuit-breaker  closed / open / half-open
    retry        exponential backoff with full jitter
    range        HTTP Range parser + Content-Range formatter
    tus          TUS v1.0.0 (Create / Patch / metadata / checksum)

  Crypto
    address      wallet-style content-addressable ids (hanzo: / mm:)
    mmpke01      multi-recipient envelope parser/builder + magic-byte sniff

  Inference (packages/memory/inference/)
    index        InferenceBackend trait + registry
    slug         provider-qualified routing (provider:model:tag)
    capabilities capability flags + task-based selector
    runtime-config hot-swap config (db_override β†’ env β†’ default)
    mock         deterministic backend (for tests)
    few-shot     ICL prompt builder (3-5 examples capped)
    refinement   Self-Refine / ReAct / Reflexion patterns
    link-types   rule + LLM link classifier (11 canonical types)
    vision       VisionBackend trait + Ollama default
    transcription Whisper-compatible HTTP client
    diarization  pyannote-compatible HTTP client + alignSpeakers / relabel
    gliner       GLiNER zero-shot NER HTTP client

  Extraction (packages/memory/extract/)
    index        registry + dispatcher
    text         native UTF-8 passthrough
    email        RFC 2822 adapter
    spreadsheet  CSV/TSV pure-JS; xlsx via registerXlsxDecoder
    archive      ZIP central-directory pure-JS reader
    exif-adapter EXIF for images
    code         tree-sitter-style syntactic chunker
    pdf          PdfSidecar HTTP adapter factory
    image        VisionBackend-backed adapter factory
    audio        TranscriptionBackend-backed adapter factory
    video        keyframe + scene + transcript adapter factory
    3d           multi-view 3D model adapter factory
    media-optimize ffmpeg variants (faststart / web-compat / 720p)
                   adapter factory
    structured   JSON-schema structured-extract adapter factory
    summarizer   fast/standard cascade adapter factory
    sprite-adapter thumbnail sprite-sheet adapter factory

Cross-runtime ports
  hanzo-memory (Python)      53 tests, ports of fusion / rerank / dedup /
                             script / fts / embed / temporal / captions /
                             tokenizer / eval / spatial / range / address /
                             graph / doc-types / circuit / retry / slug /
                             runtime-config / link-types
  bot-go (Go)                58 tests, same algorithm set ported with
                             standard-library-only primitives
  hanzoai/mcp::brain (Rust)  38 tests, same algorithm set; sha2 added as
                             dependency (blake3 stand-in)

PARITY.md is now 100% green (βœ… in-monorepo + 🟒 sibling-repo). No πŸ”΅
placeholders remain.
2026-05-11 11:44:08 -07:00

59 lines
2.1 KiB
TypeScript

/**
* Archive listing β€” ZIP/TAR/GZ.
*
* Pure-JS minimal parser: emits the file manifest. Full extraction
* lives in `hanzo-tools-fs`; the brain just needs file names + sizes
* for indexing.
*/
import type { ExtractionAdapter } from "./index.js";
export const archiveAdapter: ExtractionAdapter = {
strategy: "archive",
async extract({ bytes, filename }) {
if (isZip(bytes)) {
const entries = listZip(bytes);
const text = entries.map((e) => `- ${e.name} (${e.size} bytes)`).join("\n");
return { text: `# ${filename}\n\n${text}`, metadata: { entries } };
}
return { text: `(archive: ${filename}; unsupported format)` };
},
};
interface ZipEntry { name: string; size: number; offset: number; }
function isZip(bytes: Uint8Array): boolean {
return bytes.length >= 4 && bytes[0] === 0x50 && bytes[1] === 0x4b;
}
/**
* Walk the End-of-Central-Directory record and yield central-directory
* entries (name + size).
*/
function listZip(bytes: Uint8Array): ZipEntry[] {
const dv = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);
// EOCD signature = 0x06054b50, scan from the end.
let eocd = -1;
for (let i = bytes.length - 22; i >= 0 && i >= bytes.length - 65557; i--) {
if (dv.getUint32(i, true) === 0x06054b50) { eocd = i; break; }
}
if (eocd === -1) return [];
const cdEntries = dv.getUint16(eocd + 10, true);
const cdStart = dv.getUint32(eocd + 16, true);
let p = cdStart;
const out: ZipEntry[] = [];
for (let i = 0; i < cdEntries && p + 46 < bytes.length; i++) {
if (dv.getUint32(p, true) !== 0x02014b50) break;
const compressedSize = dv.getUint32(p + 20, true);
const uncompressedSize = dv.getUint32(p + 24, true);
const nameLen = dv.getUint16(p + 28, true);
const extraLen = dv.getUint16(p + 30, true);
const commentLen = dv.getUint16(p + 32, true);
const localHeaderOff = dv.getUint32(p + 42, true);
const name = new TextDecoder().decode(bytes.subarray(p + 46, p + 46 + nameLen));
out.push({ name, size: uncompressedSize, offset: localHeaderOff });
p += 46 + nameLen + extraLen + commentLen;
void compressedSize;
}
return out;
}