Spark ERP

AI · RAG

Answers from your knowledge, with the source shown.

A pipeline checks the question, optionally rewrites it, finds the best passages, optionally re-ranks them, and writes an answer that cites them. Every step is visible and configurable, and if nothing relevant is found it says so instead of guessing.

RAG pipelines in Spark ERP Your ERP dataModel providerPermissions Your ERP dataModel providerPermissions Human approvalProposes Audit trailAsk or trigger
RAG pipelines and its guardrails
Your ERP dataModel providerPermissionsAudit trailHuman approvalRAG pipelines
How RAG pipelines works with you
Ask or trigger

A user request or event

1
Grounded

Uses your own records

2
Proposes

A draft, not a silent change

3
You approve

Within your permissions

4
Logged

Recorded for audit

5

What happens to a question

  1. Check the question for personal data and prompt injection, and flag, mask or block.

  2. Rewrite a follow-up such as "and for my second child?" into a standalone search (optional).

  3. Find passages in the knowledge bases you chose, by meaning, by words or both.

  4. Re-rank the passages with a model (optional).

  5. Fit the passages into a size limit and write the answer with numbered citations.

  6. Check the answer, and return your no-answer message if nothing relevant was found.

The real screens

A RAG pipeline run that answers a leave question with citations to the leave policy, notice period and travel policy.
A real answer with its citations.
The step-by-step report of a pipeline run showing what each step did and how long it took.
What every step did, and how long it took.

Released like a prompt

A pipeline has immutable versions and labels such as prod, and can be called as a model by name from any application, with the citations returned alongside the answer.

See a pipeline answer from your policies.

Bring the documents and the questions.