System design

Architecture

DHARMA separates presentation, API, and RAG core — with intent routing, metadata lookup, hybrid retrieval, and multi-turn memory.

Request flow

User

Next.js

FastAPI

Intent + Metadata

Hybrid Retrieval

Groq LLM

Response

System overview

Browser

Next.js · Markdown · Sessions

API

FastAPI · Pydantic

Pipeline

Intent · Retriever · Memory

Database

PostgreSQL · pgvector

End-to-end request flow

From browser click to cited, structured wisdom response.

1User submits question in Next.js /chat (optional session_id for memory)
2POST /api/v1/chat with JSON body
3FastAPI validates input and invokes VedicWisdomPipeline
4Response mapped to typed JSON (answer, sources, related_questions)
5Frontend renders Markdown, related chips, sources, and verse card

Retrieval flow

1. QueryPreprocessor cleans query and detects scripture preference
2. MetadataRetriever resolves explicit chapter/verse references
3. Semantic search via pgvector (BGE embeddings, normalized)
4. BM25 keyword scoring on translation + explanation
5. Score-based hybrid fusion (0.65 semantic + 0.35 BM25)
6. Cross-encoder reranking and chapter deduplication
7. Top verses passed to generator (1 primary + 2 supporting)

Embedding & database flow

1. Verses loaded from CSV into PostgreSQL
2. BAAI/bge-small-en-v1.5 encodes translation + explanation
3. 384-dim normalized vectors stored in pgvector column
4. IVFFlat index built after bulk insert
5. Retriever queries index at inference time

LLM generation flow

1. Intent router selects dynamic Markdown template
2. Conversation context injected for multi-turn queries
3. Groq llama-3.3-70b-versatile generates structured answer
4. Inline citations and related follow-up questions parsed
5. Primary verse, sources, and session_id returned

System layers

Presentation — Next.js

App Router frontend with TypeScript, Tailwind CSS, Markdown rendering, related-question chips, and multi-turn session handling.

API — FastAPI

REST layer exposing health, readiness, chat (with session_id), corpus stats, and evaluation summary. Thin wrapper over the RAG pipeline.

RAG Pipeline — Python

VedicWisdomPipeline: query preprocessing, intent routing, metadata lookup, hybrid retrieval with reranking, conversation memory, and Groq generation.

Data — PostgreSQL + pgvector

867 verses with 384-dimensional BGE embeddings (bge-small-en-v1.5). IVFFlat index for approximate nearest-neighbor search.

Technology stack

Next.js 15

FastAPI

PostgreSQL

pgvector

BGE embeddings

BM25

Cross-encoder

Groq

System design

Architecture

DHARMA separates presentation, API, and RAG core — with intent routing, metadata lookup, hybrid retrieval, and multi-turn memory.

Request flow

User

Next.js

FastAPI

Intent + Metadata

Hybrid Retrieval

Groq LLM

Response

System overview

Browser

Next.js · Markdown · Sessions

API

FastAPI · Pydantic

Pipeline

Intent · Retriever · Memory

Database

PostgreSQL · pgvector

End-to-end request flow

From browser click to cited, structured wisdom response.

1User submits question in Next.js /chat (optional session_id for memory)
2POST /api/v1/chat with JSON body
3FastAPI validates input and invokes VedicWisdomPipeline
4Response mapped to typed JSON (answer, sources, related_questions)
5Frontend renders Markdown, related chips, sources, and verse card

Retrieval flow

1. QueryPreprocessor cleans query and detects scripture preference
2. MetadataRetriever resolves explicit chapter/verse references
3. Semantic search via pgvector (BGE embeddings, normalized)
4. BM25 keyword scoring on translation + explanation
5. Score-based hybrid fusion (0.65 semantic + 0.35 BM25)
6. Cross-encoder reranking and chapter deduplication
7. Top verses passed to generator (1 primary + 2 supporting)

Embedding & database flow

1. Verses loaded from CSV into PostgreSQL
2. BAAI/bge-small-en-v1.5 encodes translation + explanation
3. 384-dim normalized vectors stored in pgvector column
4. IVFFlat index built after bulk insert
5. Retriever queries index at inference time

LLM generation flow

1. Intent router selects dynamic Markdown template
2. Conversation context injected for multi-turn queries
3. Groq llama-3.3-70b-versatile generates structured answer
4. Inline citations and related follow-up questions parsed
5. Primary verse, sources, and session_id returned

System layers

Presentation — Next.js

App Router frontend with TypeScript, Tailwind CSS, Markdown rendering, related-question chips, and multi-turn session handling.

API — FastAPI

REST layer exposing health, readiness, chat (with session_id), corpus stats, and evaluation summary. Thin wrapper over the RAG pipeline.

RAG Pipeline — Python

VedicWisdomPipeline: query preprocessing, intent routing, metadata lookup, hybrid retrieval with reranking, conversation memory, and Groq generation.

Data — PostgreSQL + pgvector

867 verses with 384-dimensional BGE embeddings (bge-small-en-v1.5). IVFFlat index for approximate nearest-neighbor search.

Technology stack

Next.js 15

FastAPI

PostgreSQL

pgvector

BGE embeddings

BM25

Cross-encoder

Groq