micro/rag
Pipeline onlineConcept & docs
Retrieval labGrounded by design

Answers that show
their work.

Explore a compact RAG pipeline that retrieves from a known set of documents, validates every citation, and refuses to improvise.

01RetrieveFind relevant passages
02AnswerUse only that context
03VerifyValidate every citation
01 Ask the knowledge base
Enter to ask · Shift + Enter for a new line0/1000
Try a prompt
02 Evidence-backed response
Ready for a query

Your answer will arrive with receipts.

Ask a question to see the selected passages, similarity scores, and a complete pipeline trace.

ScopedAnswers only from indexed text
AuditableEvery source stays inspectable
CalibratedNo evidence means no answer
03 / Source library

Inspect the knowledge base.

These are the complete documents indexed by the browser demo. Compare any answer against the original wording.

data-retention.md
142 words

Data retention policy

Sample document written for the micro-rag-engine demo. The organisation described is generic.

02Customer records

Customer account records are kept for as long as the account is active and for 24 months after closure. After that period they are deleted or irreversibly anonymised.

04Support conversations

Support tickets and chat transcripts are retained for 12 months after the ticket is closed. Attachments containing personal data are deleted after 90 days unless the customer asks for them to be kept.

06Backups

Database backups are retained for 35 days. Deleted customer data can therefore remain in backups for up to 35 days after deletion, and is not restored except during a disaster recovery event.

08Deletion requests

A verified deletion request is completed within 30 days. The privacy officer records the request, the systems affected, and the completion date.

04 / About

Questions, answered plainly.

Built by Ruel T. Abion, an independent AI and cloud systems architect. Request an architecture assessment.

What is the Micro-RAG Knowledge Engine?
A small retrieval-augmented generation (RAG) pipeline that answers questions only from a known set of documents, attaches numbered citations to every answer, and refuses to answer when the documents do not contain the evidence.
Does this demo use an AI model or send my question anywhere?
No. The demo runs entirely in your browser on five bundled sample documents, using an offline keyword-overlap retriever and an extractive generator that quotes passages. There is no language model, database or API call when you ask a question. The full application can use pgvector and the Claude API.
How are citations validated?
Each citation must refer to a passage that was actually supplied to the generator. Answers without a valid citation are flagged as unverified. This proves a citation points at retrieved text; it does not prove the passage supports every claim, so the demo links each citation to the complete source document for you to check.
What happens when the documents cannot answer a question?
When no passage is relevant enough, the pipeline refuses before generating anything. A limitation: the offline keyword retriever cannot always separate an off-topic question from an in-scope one, so a semantic embedding model is recommended for production use.
Who built it?
Ruel T. Abion, an independent AI and cloud systems architect. Portfolio and contact: ruelabion.com.
Is the source code available?
The code is not public yet. A public concept and architecture repository (documentation only) is at github.com/ruelabion/micro-rag-concept. To request the code behind this demo, use the contact form at ruelabion.com/#request-code, which preselects the code-request option.
Can Ruel help build a grounded AI system for my organisation?
Yes. Ruel T. Abion offers independent architecture assessments for retrieval, AI and cloud systems. You can request one at ruelabion.com/#assessment.

ruelabion.comConcept repositoryGitHubLinkedIn