Anomaly DetectionWeld InspectionSurface InspectionVideo QC/QAAI Assistant
HOME/PRODUCTS/LINE ASSIST
LINE ASSIST · Enterprise AI Assistant (LLM)

All company knowledge becomes
a personal assistant for every employee

LINE ASSIST is a RAG-based AI assistant connected to your policies, accounts, HR and finance data. It answers only within each employee’s role permissions, cites sources on every answer — and runs fully on-premise, so company data never leaves.

Seconds
to find internal information
0
external data transfers
RBAC
role & department isolation
GATEWAY

Auth & permission gateway

SSO login with role/department (RBAC) filters applied

RETRIEVE

Hybrid search + GraphRAG

Vector + keyword fusion, expanded through relationship graphs

GENERATE

Grounded LLM generation

Answers generated only from retrieved documents

ANSWER

Cited answers

Source links on every answer · full audit logging

/01 How it works

Find it, ground it, and answer only within permission

Enterprise assistants succeed or fail on retrieval and control, not model size. LINE ASSIST retrieves precisely with a modern RAG stack, answers only what the user is allowed to see, and leaves evidence on every response.

01 · AUTH

Authenticate & filter

After AD/SSO login, role and department permissions constrain the search space itself — out-of-permission documents are never even retrieved

02 · SEARCH

Hybrid retrieval

Semantic vector search fused (RRF) with keyword search that nails exact identifiers — meaning and precision together

03 · REASON

GraphRAG & decomposition

Expansion along account–contact–contract graphs; compound questions decomposed into sub-queries (agentic)

04 · CITE

Grounded generation

Answers generated strictly from retrieved documents with mandatory source links; no evidence, no answer

* RAG instead of fine-tuning: sensitive data is never absorbed into model weights — deletion requests take effect instantly by removing the index.

HYBRID SEARCH

Exact names + semantics

Vector search alone misses identifiers like "Contract T-2024-088". Keyword fusion (RRF) captures exact matches and meaning at the same time.

GRAPHRAG

Answers connected by relationships

Accounts, contacts, contracts and receivables link as a graph — one question about an account assembles scattered facts into a single answer.

AGENTIC

Compound-question decomposition

"Who is on leave next week, and what is Account A’s receivable?" — mixed asks are split into sub-queries, retrieved separately, then synthesized.

SECURITY

Triple-layer security

Encryption at rest (AES-256) · in transit (TLS 1.3) · permission isolation (RBAC) — plus an on-premise LLM that eliminates external transfer entirely.

/02 Use Cases

Use Case Analysis

Different questions from every team — one assistant, answering within each person’s permissions.

Sales · Account management

A 360° account view in one question

Challenge

Account data lived across ERP, spreadsheets and email; preparing for a meeting took each rep 30+ minutes.

Approach

ERP records, contracts and contacts were linked as a graph, so one account name returns contracts, contacts and receivables — with sources.

Result

Meeting prep collapsed to a single question, and account history survives handovers and absences intact.

HR · General affairs

Policy questions become self-service

Challenge

Repetitive questions on leave carryover, travel expenses and benefits consumed HR hours — with inconsistent answers.

Approach

Company policies and notices were indexed, with clause-level source links mandatory on every answer, creating a self-service desk.

Result

Repetitive inquiries dropped sharply and answers unified at clause level; most onboarding questions are now handled by the assistant.

Quality · Production

Meeting decisions, retrievable again

Challenge

Decisions and owners lived in meeting minutes, but finding "what did we decide back then?" weeks later took ages.

Approach

Minutes and quality documents were indexed for decision/owner/due-date-centric queries.

Result

Past decisions surface in seconds with their source documents — meetings shifted from re-debating to progress checks.

* Figures and scenarios reflect representative deployments and may vary by process and data conditions.

/03 Specifications

Specs & Integration

Deployment
Fully on-premise (from a single GPU server), air-gapped operation supported
Data connectors
Groupware · ERP · file servers (documents/spreadsheets) · databases · messengers
Auth & permissions
AD/SSO integration, role/department RBAC — out-of-permission data blocked at retrieval
Retrieval engine
Hybrid (vector + BM25, RRF fusion) · GraphRAG · agentic query decomposition
Answer policy
Mandatory source citations · no-evidence-no-answer (hallucination control) · full query audit logs
Security
AES-256 at rest · TLS 1.3 in transit · zero external API calls (embedded LLM)
Interfaces
Web · company messenger bot; per-department knowledge bases supported
/04 FAQ

Frequently Asked Questions

Does company data go to an external AI?
No. The entire system, including the LLM, is installed on your servers with zero external API calls. During deployment we walk through the network diagram so data flow is fully transparent.
What if the AI makes things up (hallucinates)?
Answers are generated strictly from retrieved internal documents, with citations required. When no evidence is found, it is designed to say "not confirmed" instead of guessing.
What if someone asks about data beyond their permission?
Permission filtering happens at retrieval, not at answering. Out-of-permission documents are never retrieved in the first place, so they cannot leak into answers. Every query is audit-logged.
How long does deployment take?
It depends on data scope, but a phase-one build centered on policies and documents typically takes 4–6 weeks. We recommend starting with one department’s knowledge base, then expanding company-wide.
Get Started

See a demo on your own company documents

Tell us about your process and data environment — we will come back with a feasibility summary and expected impact.