Skip to content

Ask your documents. Get answers with the evidence attached.

RAG‑4i answers questions from the documents in your workspace, citing the passage behind every claim. When the documents don't cover a question, it says so.

Accounts are provisioned by your workspace administrator.

Example Answer Surface

What should we confirm before finalising a client's depreciation schedule?

Supported8 retrieved · 2 cited

Confirm the asset register reconciles with the fixed-asset ledger before the schedule is finalised.

For assets acquired during the year, depreciation runs from the date the asset was first put to use, not the invoice date.

Supporting evidence2 sources · 2 documents
  1. 1

    Year-End Procedures.pdf· p. 22

    Do not finalise a depreciation schedule until the asset register and the fixed-asset ledger reconcile, line by line, for every class of asset.
  2. 2

    Fixed Assets Advisory Note.pdf· p. 9

    For additions during the year, compute depreciation from the date the asset was first put to use; the invoice date is not the acquisition date for this purpose.

An illustration of the answer surface. Click markers to trace citations to source passages.

Every answer shows where it came from.

Citations open the exact passage a claim was drawn from, with its document and page. You decide whether the evidence holds before you rely on the answer.

And when the retrieved passages don't tell the whole story, RAG‑4i labels the answer instead of smoothing over the gap.

Year-End Procedures.pdf· p. 22
Citation [1]
“Do not finalise a depreciation schedule until the asset register and the fixed-asset ledger reconcile, line by line, for every class of asset.”

Cited by: “Confirm the asset register reconciles with the fixed-asset ledger before the schedule is finalised.”

Fixed Assets Advisory Note.pdf· p. 9
Citation [2]
“For additions during the year, compute depreciation from the date the asset was first put to use; the invoice date is not the acquisition date for this purpose.”

Cited by: “Depreciation runs from the date first put to use, not the invoice date.”

Verbatim passage fragments from workspace documents, anchored to cited claims.

Supported
Stated directly in the retrieved sources.
Partly supported
Some of this appears in the sources; the rest does not.
Sources disagree
Retrieved sources contradict each other on this point.
Not enough evidence
The documents do not cover this question.
SYSTEMS ARCHITECTURE & VERIFIED BENCHMARKS

Engineered for zero hallucination. Operating at zero GPU overhead.

RAG‑4i is architected in pure TypeScript on Supabase Edge Functions (Deno) and PostgreSQL with pgvector. Every metric below was empirically measured against frozen enterprise test suites — nothing is estimated.

Frozen Test Suite Scorecard (SciFact-30 & Statutory Corpus)Verified Production Suite
Grounding Entailment
99.0%
Checkable Claims Supported
95 / 96 claims verified against gold human truth
Zero Hallucination
0.000
Contradiction Rate
0 / 106 contradictions observed on frozen suites
Fail-Closed Gate
100%
Honest Refusal Precision
7 / 7 out-of-domain queries refused with 0 invented facts
Hybrid Search
96.7%
Recall@8 on Regulations
RRF-60 hybrid retrieval across 899 statutory chunks
Zero-GPU Edge
<120 MB
Peak Memory Footprint
Deno 2.x V8 isolates; zero OS/Python compilation binaries
Security Audit
9.5 / 10
PentAGI Security Rating
Certified production-ready: RLS isolation & prompt defense

Core Architectural Invariants

Hardened systems engineering decisions designed to scale under strict enterprise memory and security constraints.

PILLAR 01

Zero-GPU Serverless Ingestion

<120MB RAM · $0.00 Idle

Runs pure TypeScript unpdf inside Supabase Edge (Deno 2.x). Page-boundary chunking (1,000 chars / 200 overlap) permanently anchors every claim to its statutory legal page provenance without heavy PyMuPDF/Poppler dependencies.

PILLAR 02

Hybrid Retrieval with RRF-60

Dense + Sparse FTS

Fuses dense semantic search (Jina v5 embeddings via pgvector HNSW) and sparse lexical search (PostgreSQL FTS GIN) via Reciprocal Rank Fusion (k=60). Repetitive running headers lose weight in BM25 IDF without wasting LLM tokens stripping them.

PILLAR 03

In-Index Traversal (Iterative Scan)

pgvector 0.8.0+

Eliminates multi-tenant filter starvation under Row-Level Security (RLS) by configuring hnsw.iterative_scan = relaxed_order, guaranteeing full candidate retrieval even when strict workspace partition boundaries are enforced.

PILLAR 04

Pre-Inference Gating (verifyEvidence)

<150ms Fast-Path

H1 router bypasses vector search and LLMs on conversational turns. A deterministic evidence gate checks retrieved passage relevance before calling AWS Bedrock, saving ~4,500 tokens and 400ms while guaranteeing fail-closed refusals on information gaps.

Inspect Source Code & Architectural Decisions

Review full test suites, decisions record (D1–D83), and grey-box penetration reports.

From a folder of PDFs to a checked answer.

Four steps, no configuration. The same pipeline runs for every question, and it either finds the evidence or tells you it did not.

  1. Add your documents

    Upload the PDFs your team works from. Each document is parsed page by page and indexed inside your workspace.

  2. Ask in plain language

    Type the question the way you would ask a colleague. No query syntax, no keywords to guess.

  3. RAG-4i searches before it answers

    Every question retrieves passages from your workspace's documents, by meaning and by term, and the retrieved evidence is checked before any wording is produced.

  4. Read the answer, open its evidence

    Answers arrive with citations attached. Open any marker to read the passage behind it, or get a plain refusal when the documents do not cover the question.

Year-end close

Answering from 3 of 5 sources included

  • Year-End Procedures.pdf
  • Fixed Assets Advisory Note.pdf
  • Client Onboarding Checklist.pdf
  • Firm Newsletter — March.pdf
  • Archive — Old Templates.pdf

Keep answers inside the matter.

A Space is a set of sources you choose. Ask inside it and retrieval stays inside it, so the answer cannot drift into documents the matter does not touch.

Switching a source on or off changes the scope of every following question; nothing else moves.

Built for work you cannot email around.

Client documents deserve more than a shared drive. These are the boundaries the product is built on, stated as they are implemented.

Workspace isolation in the database
Documents, passages and conversations carry the workspace they belong to, and row-level security checks membership in the database on every read.
Private document storage
Files live in a private, workspace-scoped store. There are no public document URLs.
Models stay on the server
Retrieval and generation run server-side. The browser never holds a model provider key, and it never calls a provider directly.
Access follows membership
Only members of a workspace can reach its documents, and the check runs on every request. Removing someone from a workspace takes effect immediately.

Tax & advisory

  • Year-End Procedures.pdf
  • Fixed Assets Advisory Note.pdf

Audit

  • Engagement Letters.pdf
  • Sampling Guide.pdf
Row-level security: every document, passage and conversation is scoped to its workspace in the database, on every read.

Put it to work on your documents.

Sign in to open your workspace: upload a document, ask the first question, and read the answer with its evidence.

Sign in to your workspace

Accounts are provisioned by your workspace administrator.