Verifiable
Every claim is re-checked against the retrieved evidence by an independent pass. Unsupported statements are removed, not smoothed over.
Secure multilingual document intelligence with verifiable, page-level evidence.
Ask a question in your language. Groundline plans the research, searches across your documents and languages, verifies every claim, and returns an answer with the exact page and excerpt behind it.
100+ languages
- English
- العربية
- اردو
- हिन्दी
- Türkçe
- Français
- Español
- Deutsch
- 中文
- Português
- Bahasa
- Русский
Answers you can defend
Retrieving a few similar passages and letting a model fill in the gaps is fine for casual search, and a liability for decisions that have to hold up. Groundline is built for the second case.
Multilingual
Ask and answer across 100+ languages, with cross-language retrieval and automatic right-to-left rendering.
Auditable
Each answer shows its verification status, the source document, the page number, and the exact supporting excerpt.
Isolated
Customer data is tenant-scoped and encrypted, and agents never execute instructions hidden inside a document.
How it works
Four agents with narrow responsibilities. Document content is treated as untrusted data throughout, never as instructions the system will follow.
Plan
Detect the language, classify intent, and break a multi-part question into subqueries.
Retrieve
Run dense and Unicode-aware lexical search across languages for every subquery, then fuse and rerank.
Synthesize
Answer each part using only the retrieved evidence, with structured output and calculation awareness.
Verify
An independent agent re-checks every claim, calculation, and citation, and drops anything unsupported.
An answer is marked grounded only when verification passes and at least one supporting chunk survives validation. Agentic verification improves auditability, not certainty — high-impact decisions should still be confirmed against the cited pages.
Capabilities
BGE-M3 embeddings, cross-language retrieval, Unicode-aware lexical search, language-matched answers, automatic RTL rendering.
Intent planning, up to three retrieval routes, multi-hop evidence gathering, calculation-aware synthesis, independent verification.
Model-selected source chunks, document and page metadata, inspectable excerpts, and refusal to answer when evidence is missing.
PDFs up to 50 MB — headings, paragraphs, tables, OCR fallback, page preservation, deduplication, and document versioning.
Dense vector retrieval, local lexical retrieval, reciprocal-rank fusion, and optional multilingual reranking.
OIDC with scoped RBAC, server-controlled tenant scope, isolated per-tenant namespaces, upload validation, rate limiting, audit events.
Durable background jobs, progress polling, retries, timeouts, structured errors, request IDs, readiness checks, metrics, retention policies.
Multi-document chat, searchable library, device-local conversations, secure document deletion, admin-only full reset.
A 63-question regression benchmark covering numeric recall, term recall, grounding, citation coverage, latency, and estimated cost.
Two editions
Some organizations cannot send documents to an outside AI service. Groundline runs either way, without changing how answers are produced or evidenced.
Groundline Cloud
Managed hosting on fast inference and serverless vector search. Quick to stand up, straightforward to run, and the right starting point for most teams.
- Fastest path from documents to a working workspace
- Managed model inference and vector search
- Per-tenant isolation and encrypted object storage
- Ongoing updates and operational monitoring
Groundline Private
Everything runs inside the customer's own network. The language model runs on their hardware, search runs inside their database, and files stay on their disks.
- No document data leaves the customer network
- Open-weight model served on customer infrastructure
- pgvector search inside the customer's own database
- Startup checks refuse to boot a configuration that would egress data
Built for business
Governance documentation — threat model, data flow, retention, and processing terms — ships with the system and is available under review during an engagement.
Groundline has security controls and written documentation. It does not hold an external certification or completed third-party audit, and CraftWyre does not claim one.
Authentication & access
OIDC Authorization Code with PKCE and scoped RBAC in production; local API-key principals for development.
Tenant isolation
Server-controlled tenant scope with an isolated vector namespace per customer, so one tenant cannot read another's data.
Encryption
Tenant-bound object storage with per-object envelope encryption and server-side encryption at rest.
Operational resilience
Durable jobs with retries and worker crash recovery, atomic rate limits and daily quotas shared across replicas, health and readiness endpoints, and metrics.
Data governance
PII handling, retention policies, structured audit events, and a secure, admin-only workspace reset with proof of deletion.
Quality gates
A 63-question regression benchmark runs against every build. These are the thresholds a release has to clear before it ships.
≥ 90%Numeric recall
≥ 95%Citation precision
≤ 2%Unsupported-claim rate
63Benchmark questions run per release
Technology
FrontendNext.js 16, React 19, TypeScript, Tailwind CSS 4
APIFastAPI and Pydantic
Agent modelOpen-weight LLM served through a low-latency inference provider
EmbeddingsBAAI/bge-m3 — 1024-dimensional, multilingual, long-context
Vector searchPinecone Serverless in cloud; pgvector for private deployments
Metadata & lexical indexPooled PostgreSQL in production, SQLite for local development
Source objectsTenant-bound S3 with KMS envelope encryption
Background workRedis-backed job queue with PostgreSQL leases, retries, and crash recovery
PDF extractionPyMuPDF with table detection and Tesseract OCR fallback
Groundline will also be reachable at groundline.craftwyre.com once the domain is live. Every link on this site points here in the meantime.
See it refuse to guess.
The demo takes three minutes on your own documents: a cited answer, a question the documents cannot support, and the same question in another language. Bring a document set and we will run it.