mirror of
https://github.com/zooai/brain.git
synced 2026-10-30 14:50:48 +00:00
Closed every remaining π‘ / π΅ 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 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 (β in-monorepo + π’ sibling-repo). No π΅ placeholders remain.
46 lines
1.3 KiB
TypeScript
46 lines
1.3 KiB
TypeScript
/**
|
|
* Provider-qualified model slug parsing.
|
|
*
|
|
* "qwen3:8b" β default provider (Ollama)
|
|
* "ollama:qwen3:8b" β explicit Ollama
|
|
* "openai:gpt-4o" β OpenAI
|
|
* "openrouter:anthropic/claude-sonnet-4-20250514" β OpenRouter
|
|
* "llamacpp:my-model" β llama.cpp
|
|
*
|
|
* Resolves the provider via a known-provider list β anything not in the
|
|
* list is treated as part of the model id and the default provider is
|
|
* used.
|
|
*/
|
|
|
|
export const KNOWN_PROVIDERS = new Set([
|
|
"ollama",
|
|
"openai",
|
|
"openrouter",
|
|
"llamacpp",
|
|
"anthropic",
|
|
"google",
|
|
"azure",
|
|
"groq",
|
|
"together",
|
|
"mock",
|
|
]);
|
|
|
|
export interface ParsedSlug {
|
|
provider: string;
|
|
model: string;
|
|
}
|
|
|
|
export function parseSlug(slug: string, defaultProvider: string = "ollama"): ParsedSlug {
|
|
const colon = slug.indexOf(":");
|
|
if (colon === -1) return { provider: defaultProvider, model: slug };
|
|
const head = slug.slice(0, colon);
|
|
if (KNOWN_PROVIDERS.has(head)) {
|
|
return { provider: head, model: slug.slice(colon + 1) };
|
|
}
|
|
return { provider: defaultProvider, model: slug };
|
|
}
|
|
|
|
export function formatSlug(p: ParsedSlug): string {
|
|
return `${p.provider}:${p.model}`;
|
|
}
|