VerdictIQIndian Legal AI Workbench
Engineering & Research

VerdictIQ Research Lab

The retrieval, scoring, agent and verification work behind VerdictIQ — an academic major project. We describe what is built, how it works, and only numbers we have measured.

SearchCompleted

Our own Supreme Court case index (hybrid retrieval)

Supreme Court judgments from the open AWS/eCourts dataset are split into typed passages (facts, ratio), embedded with BAAI/bge-small-en-v1.5 and stored in Qdrant alongside BM25 sparse vectors. Queries run dense and keyword search together and fuse the rankings (reciprocal rank fusion), so a search for "delay in possession of flat" still finds judgments that say "non-delivery".

What we can show10,250 judgments (1985–2025) indexed as ~41,000 passages
ExplainabilityCompleted

Measured 5-dimension similarity instead of generated scores

Each precedent is scored on factual similarity and ratio similarity (calibrated cosine against the facts and ratio passages), statutory overlap (share of your provisions the judgment applies), procedural posture and relief sought. Weights are 25/25/15/15/20 and the app shows every component, so any score can be reproduced by hand.

What we can showNo percentage on screen is produced by the language model
AgentsCompleted

JurisConsult: a tool-using agent that checks its own citations

The assistant plans and calls tools — search the case index, read a judgment, look up a statute, map IPC↔BNS — for up to five steps, then a verification pass checks every case name in the answer against the judgments it actually retrieved. Anything it cannot match is flagged as unverified instead of being presented as authority.

What we can showShows its tool trace and per-citation verification status
StatutesCompleted

IPC→BNS, CrPC→BNSS, IEA→BSA transition engine

A deterministic rule engine decides which code applies from the offence date, FIR date and stage of proceedings (commencement 1 July 2024; BNSS s.531 and BSA s.170 savings) and produces dual citations. Section mappings come from the NCRB comparative tables, not from a model.

What we can showCovered by table-driven unit tests
VerificationCompleted

Citation Verifier for drafts and filings

Paste a draft and the verifier extracts SCC/AIR/SCR/INSC citations, case names and statutory sections, then resolves each against the case index, live Indian Kanoon search and the statute database — reporting what was found, what was not, and what could not be checked.

What we can showFlags invented cases and non-existent sections (e.g. "Section 999 IPC") in tests
EvaluationIn Progress

Retrieval evaluation: keyword vs dense vs hybrid

We compare BM25-only, dense-only and hybrid retrieval on held-out queries built from the corpus, reporting Recall@k and MRR. Numbers are published in the project documentation (docs/EVALUATION.md) exactly as measured.

What we can showResults reported in docs/EVALUATION.md
Model Infrastructure

Where your data goes

Language-model steps run on hosted APIs (Groq, serving open-weight models such as gpt-oss and Qwen), so the facts you submit are sent to that provider. Search embeddings are computed on our own server; search terms go to our Qdrant index and, for live results, to Indian Kanoon. Please use anonymised facts — do not paste confidential client information.

Read the security notes