@localmode/core

Local-first AI utilities for the browser. Zero dependencies. Privacy-first.

Docs website: https://localmode.dev
UI Components: https://localmode.ai
Blocks & Apps: https://localmode.ai/blocks

Installation

# Preferred: pnpm
pnpm install @localmode/core

# Alternative: npm
npm install @localmode/core

Related Packages

Package	Description
`@localmode/transformers`	HuggingFace Transformers.js provider
`@localmode/webllm`	WebLLM provider for local LLMs
`@localmode/wllama`	GGUF model provider via llama.cpp WASM
`@localmode/litert`	Google LiteRT-LM provider (`.litertlm`, WebGPU/WASM)
`@localmode/react`	React hooks for all core functions
`@localmode/chrome-ai`	Chrome Built-in AI provider
`@localmode/langchain`	LangChain.js adapters
`@localmode/devtools`	In-app DevTools widget
`@localmode/mediapipe`	Google MediaPipe Tasks — hand/pose/face landmarks, gesture recognition, vision/audio/text
`@localmode/pdfjs`	PDF text extraction
`@localmode/dexie`	Dexie.js storage adapter
`@localmode/idb`	Minimal idb storage adapter
`@localmode/localforage`	Cross-browser storage adapter

Overview

@localmode/core is a zero-dependency package containing all functions, interfaces, types, and utilities for building local-first AI applications. Provider packages (like @localmode/transformers and @localmode/webllm) implement these interfaces with specific ML frameworks.

Features

Vector Database

HNSW index for fast approximate nearest neighbor search
IndexedDB persistence with memory fallback
Cross-tab synchronization via Web Locks and BroadcastChannel
Metadata filtering with extensible operators
Scalar quantization (SQ8) for 4x storage reduction with >95% recall
Product quantization (PQ) for 8-32x storage reduction (85-92% recall)
Storage compression (compression: { type: 'sq8' }) for 4x IndexedDB disk reduction without any search recall impact
Typed metadata with generic createVectorDB<TMetadata>() for compile-time filter safety
Optional Zod schema validation on add()/addMany()
TypedFilterQuery<TMetadata> for autocomplete on filter keys and operators

Pipelines — Docs

createPipeline() - Composable multi-step workflow builder
Type-safe step chaining with progress tracking
Pre-built step factories: embedStep, chunkStep, searchStep, rerankStep, storeStep, classifyStep, summarizeStep, generateStep
AbortSignal propagation between steps
PipelineError with step context (name, index, cause)

Inference Queue — Docs

createInferenceQueue() - Priority-based task scheduling
Configurable concurrency and priority levels (interactive > background > prefetch)
Stats events for monitoring (pending, active, completed, avgLatencyMs)
AbortSignal support for queued tasks
clear() and destroy() lifecycle methods

Embeddings — Docs

embed() - Generate embeddings for single values
embedMany() - Batch embedding with progress tracking
semanticSearch() - Embed a query and search a VectorDB; results[].text round-trips chunk text stored by ingest()
Middleware support for caching, logging, validation

Embedding Drift Detection — Docs

ModelFingerprint type tracks which model produced a collection's vectors
Automatic drift detection on VectorDB.initialize() with modelDriftDetected event
checkModelCompatibility() - Read-only compatibility check
reindexCollection() - Batch re-embedding with resumability, progress, cancellation
Inference queue integration at 'background' priority
Cross-tab locking via Web Locks API

Similarity Threshold Calibration — Docs

calibrateThreshold() - Empirical threshold from corpus embedding distribution
MODEL_THRESHOLD_PRESETS - Known-good defaults for popular embedding models
getDefaultThreshold() - Instant preset lookup
Distribution statistics (mean, median, stdDev, min, max) for observability
Configurable percentile, distance function, maxSamples, AbortSignal

Multimodal Embeddings

embedImage() - Embed images into the same vector space as text
embedManyImages() - Batch image embedding
streamEmbedManyImages() - Streaming batch image embedding with progress
MultimodalEmbeddingModel interface for CLIP/SigLIP providers
Cross-modal similarity search (text-to-image, image-to-text)

Reranking — Docs

rerank() - Document reranking for improved RAG accuracy

RAG Utilities — Docs

Text chunking (recursive, markdown, code-aware, semantic)
chunk() - Split documents into optimal chunks
semanticChunk() - Embedding-aware chunking at topic boundaries
ingest() - Chunk + embed + store in one call — positional ingest(db, documents, options?) or object ingest({ db, documents, model }) with model-driven embedding via embedMany(), AbortSignal support
TEXT_METADATA_FIELD - The shared metadata key ('_text') under which ingestion stores chunk text and semanticSearch() reads it back
createBM25() - BM25 keyword search
hybridFuse(), reciprocalRankFusion() - Hybrid search combining vector and keyword results
Document loaders (Text, JSON, CSV, HTML)

Knowledge Base Engine — Docs

createKnowledgeBaseEngine() - provider-agnostic KnowledgeBaseEngine (kind: 'core') wiring chunk (off/recursive/semantic) → embed → typed-metadata VectorDB, vector search, and grounded ask with <think>…</think> reasoning stripped
Models are injected (an EmbeddingModel + a lazy getLanguageModel()) — no provider dependency, no new core dependency
Frozen KnowledgeBaseEngine contract (RawDocument, KBSearchResult, IngestOptions, SearchOptions, AskOptions, AskResult, EngineStats, ChunkingMode, DocumentSource) implemented identically by the LangChain-adapter engine in @localmode/langchain
Re-ingest-safe store (removeDocument, clear, stats), PDF page attribution, AbortSignal honored between phases

WebGPU-Accelerated Vector Search — Docs

createGPUDistanceComputer() - GPU-accelerated batch distance computation via WGSL shaders
GPUDistanceManager - WebGPU device lifecycle, buffer pooling, pipeline caching
Cosine, euclidean, and dot product distance on GPU
enableGPU option on createVectorDB() for transparent GPU acceleration
Graceful CPU fallback when WebGPU is unavailable
Configurable batch threshold (default 64) — GPU for large batches, CPU for small

Agent Framework — Docs

createAgent() - Create reusable agents with tools, instructions, and optional memory
runAgent() - One-shot agent execution with ReAct loop
createToolRegistry() - Type-safe tool registration with Zod schemas
defineTool() - Type-inferring tool definition helper — parameters and execute(params) are checked against each other, and typed tools fit ToolDefinition[] without casts
createAgentMemory() - VectorDB-backed conversation memory with semantic retrieval
Max-step guards, loop detection, duration limits
Human-in-the-loop tool approval — flag a tool with requiresApproval: true and provide an onToolApproval callback (config-level or per-run): the loop pauses before executing the flagged tool, { approved: true } resumes it unchanged, { approved: false, reason? } skips execution and feeds a denial observation back so the model adapts; AgentStep.approval records the decision. A pending decision races the run's abortSignal (aborting rejects immediately without executing) and its wait time does not count toward maxDurationMs; a gated tool with no callback fails fast with AgentError before any model call
Works with any LanguageModel provider (WebLLM, wllama, transformers, litert, chrome-ai)

Vector Import/Export — Docs

importFrom() - Import vectors from Pinecone, ChromaDB, CSV, or JSONL
parseExternalFormat() - Parse and preview external formats without importing
exportToCSV(), exportToJSONL() - Export vector collections
convertFormat() - Standalone format-to-format conversion
Auto-detect format from content, re-embedding for text-only imports
Batch processing with progress callbacks and AbortSignal

Text Generation — Docs

streamText() - Streaming text generation with async iteration
generateText() - Complete text generation with LLMs
generateObject() - Generate typed, validated JSON objects with schema
streamObject() - Stream partial JSON objects with final validation
jsonSchema() - Convert Zod schemas for structured output
Vision/Audio/Multimodal — ContentPart union type (TextPart | ImagePart | AudioPart) for sending images and audio to multimodal models
normalizeContent(), getTextContent() - Multimodal content utilities
supportsVision flag on LanguageModel interface for feature detection
AbortSignal support for cancellation
Works with WebLLM (Phi 3.5 Vision), Transformers (Qwen3.5), wllama, and litert (Gemma 4 E2B/E4B + Qwen3 0.6B, text-only) for local LLM inference

Classification & NLP — Docs

classify(), classifyMany() - Text classification
classifyZeroShot() - Zero-shot classification with custom labels
extractEntities(), extractEntitiesMany() - Named Entity Recognition

Audio — Docs

transcribe() - Speech-to-text with Whisper / Moonshine models
synthesizeSpeech() - Text-to-speech synthesis (one-shot)
streamSynthesizeSpeech() - Streaming text-to-speech: clause-by-clause Float32Array audio for low first-audio latency
playStreamedSpeech() - Gap-free Web Audio playback for the streaming iterable, with pause / resume / stop / abort
splitIntoClauses() - Pure clause splitter (sentence + comma/semicolon boundaries, abbreviation- and decimal-aware)
Kokoro TTS via @localmode/transformers: 29 voices, speed control, phonemizer-backed synthesis
classifyAudio() - Audio classification
classifyAudioZeroShot() - Zero-shot audio classification
Word and segment-level timestamps, multi-language support

Live Transcribe — Docs

createLiveTranscriber() - Streaming microphone-driven STT with VAD
Push-to-talk and open-mic modes with auto-segmentation
Built-in EnergyVADProvider (zero-dependency, AudioWorklet-based) and SileroVADProvider adapter
Barge-in detection during external TTS playback
createTurnTaker() orchestrator for full voice loops (user → planner → agent → user)

Vision — Docs

classifyImage() - Image classification
classifyImageZeroShot() - Zero-shot image classification
captionImage() - Image captioning with BLIP/Florence models
segmentImage() - Image segmentation / background removal
detectObjects() - Object detection with bounding boxes
extractImageFeatures() - Image feature extraction for similarity
imageToImage() - Image super resolution / image-to-image transformation
estimateDepth() - Monocular depth estimation
detectHands() - 21-point hand landmark detection (MediaPipe)
detectPose() - 33-point body pose landmark detection (MediaPipe)
detectFace() - Face detection with bounding boxes and keypoints (MediaPipe)
detectFaceLandmarks() - 478-point face mesh detection (MediaPipe)
recognizeGesture() - Hand gesture recognition (MediaPipe)

Text Generation (Complete)

generateText() - Complete text generation with LLMs
wrapLanguageModel() - Language model middleware for caching, logging, guardrails
composeLanguageModelMiddleware() - Compose multiple middleware

Semantic Cache

createSemanticCache() - Cache LLM responses using embedding similarity
semanticCacheMiddleware() - Transparent caching middleware for generateText/streamText

Translation — Docs

translate() - Text translation between languages
detectLanguage() - Language detection (110 languages, ISO 639-1 codes)

Summarization — Docs

summarize() - Text summarization with configurable length

Fill-Mask — Docs

fillMask() - Masked token prediction (BERT-style)

Question Answering — Docs

answerQuestion() - Extractive question answering from context

OCR — Docs

extractText() - Optical character recognition from images

Document QA — Docs

askDocument() - Question answering on document images
askTable() - Question answering on tabular data

Storage — Docs

IndexedDBStorage - Persistent browser storage
MemoryStorage - In-memory fallback
Automatic quota management and cleanup

Tamper-evident Audit Log — Docs

createAuditLog() - Append-only, hash-chained, signed entries
HMAC-SHA-256 default; Ed25519 opt-in via injected CryptoKey
verifyChain() - Cancellable walker reports hash_mismatch / prev_hash_mismatch / signature_mismatch
Optional AES-GCM payload encryption — chain still verifies without the encryption key
exportAuditLog() - JSON Lines output for customer-controlled S3 / Azure Blob sync
deriveAuditKey() / generateEphemeralAuditKey() - PBKDF2 + ephemeral helpers
Zero-dep, no network calls — Web Crypto only

Evaluation SDK — Docs

accuracy(), precision(), recall(), f1Score() - Classification metrics
bleuScore(), rougeScore() - Text generation metrics
mrr(), ndcg() - Retrieval metrics
evalCosineDistance() - Vector quality metric
confusionMatrix() - Structured confusion matrix with helper methods
evaluateModel() - Orchestrator: run model against dataset, apply metric, get report
Custom metrics via MetricFunction type

Capabilities Detection — Docs

isWebGPUSupported() - Check WebGPU availability
detectCapabilities() - Full device capability report
Automatic fallback recommendations
computeOptimalBatchSize() - Adaptive batch sizing from device hardware (CPU cores, memory, GPU)
recommendModels() - Ranked model recommendations for any task based on device capabilities
registerModel() / getModelRegistry() - Curated model catalog with runtime extensibility

Quick Start

Semantic Search with PDF Documents

import { createVectorDB, embed, embedMany, chunk, rerank } from '@localmode/core';
import { transformers } from '@localmode/transformers';

// Create embedding model
const embeddingModel = transformers.embedding('Xenova/bge-small-en-v1.5');

// Create vector database
const db = await createVectorDB({
  name: 'documents',
  dimensions: 384,
});

// Chunk and embed document
const chunks = chunk(documentText, {
  strategy: 'recursive',
  size: 512,
  overlap: 50,
});

const { embeddings } = await embedMany({
  model: embeddingModel,
  values: chunks.map((c) => c.text),
});

// Store in database
await db.addMany(
  chunks.map((c, i) => ({
    id: `chunk-${i}`,
    vector: embeddings[i],
    metadata: { text: c.text },
  }))
);

// Search
const { embedding: queryVector } = await embed({
  model: embeddingModel,
  value: 'What is machine learning?',
});

const results = await db.search(queryVector, { k: 10 });

// Optional: Rerank for better accuracy
const rerankerModel = transformers.reranker('Xenova/ms-marco-MiniLM-L-6-v2');
const reranked = await rerank({
  model: rerankerModel,
  query: 'What is machine learning?',
  documents: results.map((r) => r.metadata.text),
  topK: 5,
});

Vector Quantization (4x Storage Reduction)

import { createVectorDB } from '@localmode/core';

// Enable scalar quantization — same API, 4x less storage
const db = await createVectorDB({
  name: 'quantized-docs',
  dimensions: 384,
  quantization: { type: 'scalar' },
});

// Usage is identical — quantization is transparent
await db.add({ id: 'doc1', vector: embedding, metadata: { text: 'Hello' } });
const results = await db.search(queryVector, { k: 10 });

// Recalibrate if vector distribution changes significantly
await db.recalibrate({ onProgress: (done, total) => console.log(`${done}/${total}`) });

LLM Chat with Streaming

import { streamText } from '@localmode/core';
import { webllm } from '@localmode/webllm';

const model = webllm.languageModel('Llama-3.2-1B-Instruct-q4f16_1-MLC');

const result = await streamText({
  model,
  prompt: 'Explain quantum computing in simple terms',
  maxTokens: 500,
});

for await (const chunk of result.stream) {
  process.stdout.write(chunk.text);
}

Core Exports

Vector Database

import {
  createVectorDB,
  createVectorDBWithWorker,
  HNSWIndex,
  cosineSimilarity,
  euclideanDistance,
  dotProduct,
} from '@localmode/core';

Embeddings

import {
  embed,
  embedMany,
  streamEmbedMany,
  semanticSearch,
  wrapEmbeddingModel,
} from '@localmode/core';

Multimodal Embeddings

import {
  embedImage,
  embedManyImages,
  streamEmbedManyImages,
} from '@localmode/core';

Reranking

import { rerank } from '@localmode/core';

RAG Utilities

import {
  // Chunking
  chunk,
  recursiveChunk,
  markdownChunk,
  codeChunk,
  // BM25
  createBM25,
  // Hybrid Search
  hybridFuse,
  reciprocalRankFusion,
  // Ingestion
  ingest,
  TEXT_METADATA_FIELD,
} from '@localmode/core';

Knowledge Base Engine

import { createKnowledgeBaseEngine } from '@localmode/core';
import type {
  KnowledgeBaseEngine,
  RawDocument,
  KBSearchResult,
  AskResult,
  EngineStats,
} from '@localmode/core';

const engine = createKnowledgeBaseEngine({
  embeddingModel, // an EmbeddingModel you construct (e.g. transformers.embedding(...))
  getLanguageModel: () => languageModel, // lazy — loaded on first ask()
});
await engine.ingest(docs, { chunking: 'recursive', chunkSize: 500 });
const hits = await engine.search('privacy and encryption', { topK: 10 });
const { answer, sources } = await engine.ask('What is encrypted?');

Text Generation

import { streamText, generateText } from '@localmode/core';

Storage

import {
  IndexedDBStorage,
  MemoryStorage,
  createStorage,
  getStorageQuota,
  requestPersistence,
  cleanup,
} from '@localmode/core';

Capabilities

import {
  detectCapabilities,
  isWebGPUSupported,
  isIndexedDBSupported,
  isAudioWorkletSupported,
  isMediaCaptureSupported,
  isLiveTranscribeSupported,
  checkModelSupport,
  getRecommendedFallbacks,
  recommendModels,
  registerModel,
  getModelRegistry,
  DEFAULT_MODEL_REGISTRY,
} from '@localmode/core';

Classification

import { classify, classifyMany, classifyZeroShot } from '@localmode/core';

NER

import { extractEntities, extractEntitiesMany } from '@localmode/core';

Audio

import {
  transcribe,
  synthesizeSpeech,
  streamSynthesizeSpeech,
  playStreamedSpeech,
  splitIntoClauses,
  DEFAULT_ABBREVIATIONS,
  classifyAudio,
  classifyAudioZeroShot,
} from '@localmode/core';

Live Transcription & VAD

import {
  createLiveTranscriber,
  createTurnTaker,
  EnergyVADProvider,
  SileroVADProvider,
  registerEnergyVADWorklet,
  createScriptProcessorVADNode,
  isLiveTranscribeSupported,
  isAudioWorkletSupported,
  isMediaCaptureSupported,
  MediaNotSupportedError,
} from '@localmode/core';

Vision

import {
  classifyImage,
  classifyImageZeroShot,
  captionImage,
  segmentImage,
  detectObjects,
  extractImageFeatures,
  imageToImage,
  estimateDepth,
  detectHands,
  detectPose,
  detectFace,
  detectFaceLandmarks,
  recognizeGesture,
  HAND_CONNECTIONS,
  POSE_CONNECTIONS,
  FACE_CONNECTIONS,
  GESTURE_CATEGORIES,
} from '@localmode/core';

Translation

import {
  translate,
  detectLanguage,
  SUPPORTED_LANGUAGES,
  getLanguageName,
} from '@localmode/core';

Summarization

import { summarize } from '@localmode/core';

Fill-Mask

import { fillMask } from '@localmode/core';

Question Answering

import { answerQuestion } from '@localmode/core';

OCR

Supports an optional prompt parameter for generative OCR models (e.g., GLM-OCR, LightOnOCR-2) to control recognition mode — table structure, formula/LaTeX, or plain text extraction.

import { extractText } from '@localmode/core';

Document QA

import { askDocument, askTable } from '@localmode/core';

Middleware

import {
  wrapEmbeddingModel,
  wrapVectorDB,
  cachingMiddleware,
  loggingMiddleware,
  retryMiddleware,
  rateLimitMiddleware,
  validationMiddleware,
  piiRedactionMiddleware,
  encryptionMiddleware,
} from '@localmode/core';

Security

import { encrypt, decrypt, deriveKey, isCryptoSupported, redactPII } from '@localmode/core';

Audit Log

import {
  createAuditLog,
  verifyChain,
  exportAuditLog,
  deriveAuditKey,
  generateEphemeralAuditKey,
  AuditLogError,
} from '@localmode/core';

Differential Privacy

import {
  // Embedding middleware — adds calibrated noise to embeddings
  dpEmbeddingMiddleware,
  // Classification middleware — randomized response for labels
  dpClassificationMiddleware,
  randomizedResponse,
  // Privacy budget tracking
  createPrivacyBudget,
  // Noise mechanisms
  gaussianNoise,
  laplacianNoise,
  // Sensitivity calibration
  calibrateSensitivity,
} from '@localmode/core';

Cross-Tab Sync

import { createBroadcaster, createLockManager, isWebLocksSupported } from '@localmode/core';

Network

import {
  getNetworkStatus,
  onNetworkChange,
  isOnline,
  isOffline,
  waitForOnline,
} from '@localmode/core';

Events

import { createEventEmitter, globalEventBus } from '@localmode/core';

Errors

import {
  LocalModeError,
  EmbeddingError,
  ModelNotFoundError,
  StorageError,
  QuotaExceededError,
  ValidationError,
  AuditLogError,
  MediaNotSupportedError,
  formatErrorForUser,
} from '@localmode/core';

Testing Utilities

import {
  createMockEmbeddingModel,
  createMockRerankerModel,
  createMockStorage,
  createMockVectorDB,
  createTestVector,
  createSeededRandom,
  createMockHandLandmarkModel,
  createMockPoseLandmarkModel,
  createMockFaceDetectionModel,
  createMockFaceLandmarkModel,
  createMockGestureRecognitionModel,
  createMockLanguageDetectionModel,
  // Framework-agnostic StorageAdapter contract suite for custom adapters —
  // returns Array<{ name, run }>; register with `for (const c of suite) it(c.name, c.run)`
  createStorageAdapterConformanceSuite,
} from '@localmode/core';

Architecture

Zero-Dependency Core

@localmode/core has zero external dependencies. All functionality is implemented using native browser APIs:

Vector Search: Custom HNSW implementation
Storage: IndexedDB + Memory fallback
Encryption: Web Crypto API
Sync: Web Locks + BroadcastChannel

Provider Pattern

Provider packages implement core interfaces:

// @localmode/transformers - HuggingFace Transformers.js
import { transformers } from '@localmode/transformers';
const embedder = transformers.embedding('Xenova/bge-small-en-v1.5');

// @localmode/webllm - WebLLM for local LLMs
import { webllm } from '@localmode/webllm';
const llm = webllm.languageModel('Llama-3.2-1B-Instruct-q4f16_1-MLC');

Function-First API

All operations are exposed as top-level async functions:

// Correct: Top-level functions
const { embedding } = await embed({ model, value: 'Hello' });

// Wrong: Class methods
const embedder = new Embedder(model);
await embedder.embed('Hello');

Options Object Pattern

All functions accept a single options object:

const result = await embed({
  model: embeddingModel,
  value: 'Hello world',
  abortSignal: controller.signal, // Optional
});

Structured Results

All functions return structured result objects:

interface EmbedResult {
  embedding: Float32Array;
  usage: { tokens: number };
  response: { modelId: string; timestamp: Date };
}

User Extensibility

Implement any core interface to create custom providers:

Custom Storage

import type { Storage } from '@localmode/core';

class MyRedisStorage implements Storage {
  async get(key: string) {
    /* ... */
  }
  async set(key: string, value: StoredDocument) {
    /* ... */
  }
  async delete(key: string) {
    /* ... */
  }
  async keys() {
    /* ... */
  }
  async clear() {
    /* ... */
  }
  async close() {
    /* ... */
  }
}

const db = await createVectorDB({
  storage: new MyRedisStorage(),
});

Custom Embedding Model

import type { EmbeddingModel } from '@localmode/core';

class MyAPIEmbedder implements EmbeddingModel {
  readonly modelId = 'custom:my-embedder';
  readonly provider = 'custom';
  readonly dimensions = 768;
  readonly maxEmbeddingsPerCall = 100;
  readonly supportsParallelCalls = true;

  async doEmbed(options: DoEmbedOptions) {
    // Your implementation
    return {
      embeddings: [new Float32Array(768)],
      usage: { tokens: 10 },
    };
  }
}

const { embedding } = await embed({
  model: new MyAPIEmbedder(),
  value: 'Hello',
});

Browser Compatibility

Browser	WebGPU	WASM	IndexedDB	Workers
Chrome 80+	113+	✅	✅	✅
Edge 80+	113+	✅	✅	✅
Firefox 75+	Nightly	✅	✅	✅
Safari 14+	18+	✅	✅	⚠️

Platform Notes

Safari/iOS: Private browsing blocks IndexedDB → use MemoryStorage
Firefox: WebGPU shipped in Firefox 141+ on Windows and 147+ on macOS Apple Silicon — falls back to WASM on Linux/older versions
SharedArrayBuffer: Requires cross-origin isolation

Privacy Guarantees

No telemetry - We don't track anything
No network requests - Core package makes zero network calls
Data stays local - All processing happens in your browser
Open source - Audit the code yourself

License

MIT

Uh oh!

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

@localmode/core

Installation

Related Packages

Overview

Features

Vector Database

Pipelines — Docs

Inference Queue — Docs

Embeddings — Docs

Embedding Drift Detection — Docs

Similarity Threshold Calibration — Docs

Multimodal Embeddings

Reranking — Docs

RAG Utilities — Docs

Knowledge Base Engine — Docs

WebGPU-Accelerated Vector Search — Docs

Agent Framework — Docs

Vector Import/Export — Docs

Text Generation — Docs

Classification & NLP — Docs

Audio — Docs

Live Transcribe — Docs

Vision — Docs

Text Generation (Complete)

Semantic Cache

Translation — Docs

Summarization — Docs

Fill-Mask — Docs

Question Answering — Docs

OCR — Docs

Document QA — Docs

Storage — Docs

Tamper-evident Audit Log — Docs

Evaluation SDK — Docs

Capabilities Detection — Docs

Quick Start

Semantic Search with PDF Documents

Vector Quantization (4x Storage Reduction)

LLM Chat with Streaming

Core Exports

Vector Database

Embeddings

Multimodal Embeddings

Reranking

RAG Utilities

Knowledge Base Engine

Text Generation

Storage

Capabilities

Classification

NER

Audio

Live Transcription & VAD

Vision

Translation

Summarization

Fill-Mask

Question Answering

OCR

Document QA

Middleware

Security

Audit Log

Differential Privacy

Cross-Tab Sync

Network

Events

Errors

Testing Utilities

Architecture

Zero-Dependency Core

Provider Pattern

Function-First API