Architecture

How Nyaya answers

Query → retrieve → cite. A four-stage pipeline over an indexed legal corpus, exposed via web, API, and MCP.

Pipeline

Click a stage to expand technical detail. Each stage is a discrete service with its own contract.

Query

User question is parsed, normalised, and embedded for retrieval.

Retrieve

MCP tools fetch matching provisions and judgments from the indexed corpus (semantic search fuses embedding retrieval + reranking).

Cite

Synthesis LLM drafts the answer with inline citations from retrieved passages.

MCP

get_section / get_article parse and fetch each reference from the corpus; MCP exposes 16 tools to editors.

Stack

Services and infrastructure. Tool tags are generic roles, not vendor endorsements.

Services

Embedding model

Converts text queries and corpus passages into dense vectors for semantic search.

nvidia/nemotron-3-embed-1b

Vector store

Stores and retrieves passage embeddings with cosine similarity search at scale.

pgvector

Reranker

Reorders top-k retrieved passages by relevance to the specific query using a cross-encoder model.

llama-nemotron-rerank-vl-1b-v2

LLM

Drafts the cited answer from retrieved context. Constrained to cite only from the retrieved set.

Nemotron-3.5 Lightning 30B

Citation parser

Parses citation strings ('IPC s.302', 'Art.21') and fetches the matching provision. Built into get_section / get_article.

nyaya/get_section + get_article

Infrastructure

Hosting

Web app + API + MCP server, served from one container. Deploy region chosen at deploy time.

Railway / Docker

Refresh jobs

Ingestion is manual via the hydration notebook. Automated scheduled crawlers are planned.

hydrate.ipynb notebook

Observability

Latency, retrieval quality, and error tracking.

[tool: TBD]

Object storage

Stores raw PDFs and parsed text from sources for audit and re-indexing.

[planned]

Search index

Full-text fallback alongside vector search for exact-match section/article lookups.

Postgres FTS

Data flow

From question to cited answer, step by step. Each step links to the pipeline stage it belongs to.

1User submits a natural-language legal question through the web chat or an MCP client.

2Supervisor LLM reasons briefly about which corpus tools to call, then emits all tool calls in one message for parallel execution.

3MCP tools (semantic_query, get_section, get_article, …) retrieve matching provisions and judgments from the indexed corpus in parallel.

4Synthesis LLM composes the final answer using only the retrieved tool results, with inline [[act: X, ref: Y]] citations.

5get_section / get_article parse each [[act, ref]] citation inline and fetch the matching provision from the corpus before the answer is displayed.

MCP server

Nyaya ships an MCP server so Claude, Cursor, and opencode can query the corpus directly from your editor. Add this config to your client:

mcp.json

{
  "mcpServers": {
    "nyaya": {
      "url":"https://nyaya.parag.tech/mcp",
      "transport":"http"
    }
  }
}

Supported MCP tools

Loading tools…

Openness & roadmap

What is open today and what is planned. Honest status. No fake repo links.

Today

MCP server (HTTP endpoint at /mcp)live

Open-source MCP server (Apache-2.0)live

Self-host recipe (Docker + docker-compose)live

Web chat + citation engine (Nemotron-3.5 Lightning)live

Corpus data dumpsclosed

Planned

Open corpus data (CC-BY)planned

Citation resolver API (public)planned

Overruled-status check for judgmentsplanned

Automated refresh jobsplanned