manish-9245/scalable-llm-search

Purpose and Entry Points

Indriya AI is a local-first luxury-jewellery search, voice, and concierge backend with an administrative AI ingestion pipeline. The Fastify application starts in server.js, serves the browser UI from public/index.html and public/app.js, and exposes search, chat, transcription, ingestion, queue-admin, and health routes.

HTTP Coordination and Runtime Boundaries

server.js is the request coordinator: it accepts text/voice interactions, invokes local search and localization paths, and dispatches privileged ingestion work. It runs in Node.js/Fastify; static UI and local ONNX/WASM inference execute alongside the service. Deployment packaging is defined by Dockerfile and Railway settings in railway.json.

Query Understanding and Local Concierge

The local parsing layer converts raw customer wording into structured filters, including typo correction, jewellery terminology, vernacular mappings, price ranges, purity/color normalization, and exclusions. Core vocabulary and parsing logic are in src/utils/terminology.js, src/utils/slang.js, and src/utils/QueryBuilder.js. Multilingual template-based conversational output is configured in src/config/translations.js; this customer-facing path is intentionally independent of cloud LLM calls.

Adaptive Hybrid Search

src/services/searchService.js executes the primary retrieval flow: parsed filters become relational, GIN-array exclusion/tag, and pgvector semantic retrieval streams, then are merged through reciprocal-rank fusion. src/services/discoveryService.js supports discovery-oriented catalog access. Embeddings use local MiniLM/ONNX inference configured through src/config/llm.js, while database access is centralized in src/config/db.js. PostgreSQL/pgvector is the durable search boundary, with schema and indexes represented by schema.sql.

Caching, Sessions, and Live Pricing

Redis provides low-latency search/product caching and BullMQ backing, configured in src/config/redis.js and src/config/queue.js. Session continuity is handled by src/utils/sessionMerge.js. src/services/rateFetcherService.js periodically fetches Indriya’s public rate fragment, derives related metal rates, persists daily values, and invalidates affected Redis keys; it is the only material product-price external data dependency.

Administrative Ingestion Pipeline

Ingestion is separated from customer runtime. src/services/ingestionService.js submits or processes background catalog enrichment; src/mastra/agent.js and src/mastra/tools.js define the Mastra/Gemini-backed multimodal analysis boundary. Gemini is used only to create structured product dossiers, which are normalized and stored in PostgreSQL. Redis/BullMQ decouples upload/request handling from this potentially slow, retryable work.

Voice, Initialization, and Operations

The transcription route uses local Whisper ONNX/WASM with language routing; model assets are prepared during build/startup through src/config/pre_cache.js. Database initialization and ontology seeding begin at db_init.mjs, with base schema in schema.sql. Structured logging is provided by src/utils/logger.js; queue visibility is mounted by the server. Required infrastructure is PostgreSQL 16+ with pgvector, Redis 7+, and a Gemini API key solely for ingestion.