Signal & Gate ← Architecture
03 · Dataset & Run It

What the corpus actually looks like, and how to run it yourself.

Real documents from four categories, a live chunking preview for your own files, and the exact commands to bring the full system up on your machine.

Example Documents

Eleven documents, four categories

A cross-section of the actual corpus. Teal tags are verbatim from data/corpus/; amber tags are illustrative examples built in the same schema, standing in for client and support records that can't be shown here (see the note below).

Client-record and support-ticket examples here use fictional companies (Cedar Retail, Northwind Grocery) built in the same schema as the real files. The shipped corpus does include real client and support records, but they name actual companies with fabricated financial details attached, so they're excluded from this public page. Everything tagged "real corpus" above is verbatim from the repository.

Try It Yourself

This page can show you the pipeline. Only your own copy can run it.

There's no LLM behind this static page. The answers on the Overview page are scripted. Everything below is real: a real sample dataset, the real chunking algorithm running in your browser, and the exact commands to bring the full agent up on your machine.

Download the sample corpus

Three files, safe to demo publicly: two verbatim from data/corpus/, one illustrative.

basic_company_info.json8 docs · real corpus · offices, leadership, HR policy
payment_processor_analysis.json11 docs · real corpus · Razorpay / Stripe / PayU pricing
illustrative_client_examples.json4 docs · fictional · client + support record schema

Saved as .json, rename to .jsonl and drop into data/corpus/ before running the ingest script.

Bring your own data

Upload a .txt, .md or .jsonl file and see exactly how chunk_document() would slice it: 200-word windows, 40-word overlap, title prepended to the first chunk. Runs entirely in your browser; nothing leaves this page.

Run the full system locally

Needs your own OpenAI API key. Redis and both agent + dashboard containers come up with one command.

# 1. clone and configure git clone https://github.com/vrhsr/veritas-agent.git cd veritas-agent cp .env.example .env # add your OPENAI_API_KEY # 2. build the retrieval indexes python scripts/ingest_corpus.py # 3. bring the stack up docker compose -f config/docker-compose.yml up

API on localhost:8000 · dashboard on localhost:8501.

Built On

The stack underneath

Orchestration

LangGraphLangChain

LLM

gpt-4o-minigpt-4o (prod)

Retrieval

FAISSBM25OkapiMiniLM-L6-v2Cross-Encoder

Memory

RedisPer-user FAISS

Serving

FastAPISSEPrometheus

Observability

structlogtiktokenLangSmith

Dashboard / Eval

StreamlitPlotlyRAGAS

Deployment

Docker ComposeRender.com