Documentation
Meridian Intel
Everything you need to upload, query, and understand results from your documents — zero hallucination, every domain.
Introduction
What Meridian Intel does and how it's different.
Meridian Intel is a Retrieval-Augmented Generation (RAG) platform built for accuracy-critical document review. Upload a document, ask a question in plain language, and get an answer grounded entirely in what the document says — never beyond it.
If no relevant content is found in your document, the platform tells you directly rather than generating a guess. The AI model is never called when retrieval returns nothing.
Answers are extracted verbatim from your document by default — exact words, not paraphrases. Every fact is cited with its source page and chunk reference. If you need a simpler explanation, ask explicitly (e.g. "explain this in simpler terms") and the system will paraphrase while still grounding every claim in the document.
Built by Mazal Arc, Meridian Intel was designed first for pharmaceutical compliance review — Certificates of Quality, BSE/TSE declarations, Chain of Custody documentation — and now supports 14+ domains including legal, finance, research, and IT.
Quick Start
Three steps from zero to your first answer.
↑ Upload in the workspace top bar. Select a PDF, TXT, DOCX, or DOC file under 50MB. Previously uploaded documents can be loaded instantly from History — no re-upload needed.Domain Fields
14+ supported domains across 5 categories. Unrecognized fields fall back to General Purpose without failing.
Science & Research
Business & Legal
Technology
Custom
Don't see your field? Type a custom field name during upload. The chunking and retrieval engine works the same — only the PII scrubbing defaults differ from pre-configured fields.
Supported Formats
What you can upload today, and what's coming.
| Format | Extraction method | Status |
|---|---|---|
| pdfplumber text extraction | Available | |
| .txt | Plain text ingestion | Available |
| .docx / .doc | python-docx + table extraction | Available |
| Scanned PDF | OCR (EasyOCR — selected, integration in v2.1) | v2.1 |
| .csv | Tabular chunking | v2.2 |
| .png / .jpg | VLM diagram analysis | v2.2 |
If you upload a PDF with no extractable text (a scanned image), the platform detects this automatically and lets you know OCR support is coming rather than returning an empty or broken result.
Tier Guide
Five tiers, each with distinct retrieval precision and feature access.
| Tier | Chunk size | Results | Key unlock |
|---|---|---|---|
| Ephemeral | 600 tokens | 3 | Free, session-only, 1 domain |
| Segment | 512 tokens | 5 | 5 domains, document history |
| Lattice | 384 tokens | 6 | 10 domains, eval scores, full telemetry |
| Vector | 256 tokens | 8 | All 14+ domains, PDF highlights, API key |
| Enclave | 200 tokens | 10 | Per-user isolation, audit trail, custom SLA |
Smaller chunk sizes on higher tiers mean more precise retrieval — the platform searches more granular sections of your document, which matters most on dense regulatory or technical content.
Understanding Results
How to read what the platform gives you back.
Eval Scores
How much of the answer's vocabulary is grounded in the retrieved document text. A score near 1.0 means the answer uses exact words and phrases from the document. Verbatim extraction mode typically achieves 0.90+ faithfulness.
How much of your query's key terms appear in the answer. Measures whether the answer addressed your actual question. Uses stemming so "work" matches "works" and "working".
What fraction of the retrieved context was used in the answer. Lower scores are normal when the retrieved chunks contain more content than needed — the answer extracts only the relevant parts.
Zero-Hit Responses
When the platform tells you it found nothing relevant, this is a feature, not a failure. It means retrieval came back empty and the AI model was never called — preventing a fabricated answer. Try rephrasing your question or confirm the document contains the information you're looking for.
Source Citations
Every answer shows the page and section reference it was drawn from (e.g. p.1 s2), visible in the audit panel alongside the exact retrieved text.
API Reference
Programmatic access for Vector and Enclave tier users.
Requesting a Key
API keys are currently issued manually during the validation phase. Go to Settings → Identity Hub → API Key and click Request API Key. This opens a pre-filled email to our support team — you'll receive your key within 24 hours.
Authentication
Authorization: Bearer sk_live_xxxxxxxxxxxxxxxx
Query Endpoint
POST https://meridianintel.app/query { "query": "Is this batch EU GMP compliant?", "field": "certificate_of_quality", "tier": "vector" }
Ingest Endpoint
POST https://meridianintel.app/ingest Content-Type: multipart/form-data file: <binary> field: "legal_contract" tier: "vector"
Per-key rate limiting via Upstash Redis is on the v2.1 roadmap. During validation, usage is monitored manually — reach out if you need higher throughput.
Security & Privacy
How your data is protected.
Authentication
Sign-in is available via Google SSO (OAuth 2.0, client-side GIS flow) or email verification with a 6-digit code. Enterprise teams on Enclave tier can use Company SSO with SAML 2.0 or OIDC. Sessions auto-lock after 15 minutes of inactivity to protect sensitive document data, with a 2-minute warning before logout.
PII Scrubbing
Nine categories of personally identifiable information are scrubbed before any content reaches an AI model — email, phone, SSN, credit card, IP address, passport, date of birth, address, and names. Pharmaceutical fields use adjusted defaults so lot numbers and catalog IDs aren't incorrectly flagged.
Inference
Primary inference runs on a locally-hosted Mistral-7B model. Your document content does not reach OpenAI, Anthropic, or any third party during normal operation. The AUTO Shield cascade falls back to Gemini 2.5 Flash Lite, then OpenAI gpt-4o-mini, only when local inference is unavailable — and only PII-scrubbed text is ever sent to any cloud provider.
local_only Mode
Enclave-tier users and compliance-sensitive deployments can enable local_only engine mode. In this mode, the platform never contacts any cloud provider under any condition — if local inference cannot complete, the platform returns an honest failure message rather than silently falling back to a cloud API. This provides a verifiable, auditable zero-cloud guarantee for government, legal, and regulated-industry use cases.
Data Isolation
Enclave tier users receive a fully isolated per-user vector store collection. Other tiers use a shared collection namespace with strict user-ID and document-ID tagging — retrieval is scoped to the specific document you uploaded, with zero cross-document contamination verified across real multi-document test scenarios.
Full details are in our Privacy Policy and Terms of Service.
FAQ
Common questions. For anything else, use the chat widget on the homepage or email us.
Does Meridian Intel train on my documents?
No. Neither Mazal Arc nor any third-party AI provider uses your uploaded documents or queries to train or fine-tune models.
What happens if I ask something the document doesn't answer?
The platform returns a clear "no relevant content found" response rather than guessing. This is the zero-hallucination guarantee in action.
Can I delete my documents?
Yes, at any time via the platform interface or by emailing support. Deletion completes within 30 days.
What's the difference between the inference engines?
Mistral-7B runs locally on our infrastructure with zero data leaving our servers. Gemini 2.5 Flash Lite and OpenAI gpt-4o-mini are cloud options, sending only PII-scrubbed content. You can manually select any engine from the overflow menu — selecting an engine disables the AUTO Shield cascade and routes directly to that engine. Toggle AUTO Shield back on to re-enable the Mistral → Gemini → OpenAI cascade. In local_only mode, cloud fallback is disabled entirely — Mistral runs exclusively, and the platform returns an honest failure rather than routing to any cloud API.
What is the local_only engine mode?
A zero-cloud inference guarantee for compliance-sensitive use cases. When enabled, your document content never leaves your infrastructure under any condition. Latency is higher than AUTO mode (dependent on your hardware), but the guarantee is complete and verifiable — no exceptions, no silent fallbacks.