AI may
- explain a treated case in plain language
- compare evidence and point to disagreements
- draft questions, briefs, or follow-up notes
- adapt a verified explanation for another audience
- suggest missing evidence for a person to retrieve
Applied enterprise RAG
Use AI to amplify meaningful work, not replace judgment.
Shakti is a small reference architecture for a serious idea. Retrieve only the evidence needed for one case. Treat it before a model sees it. Let AI explain, compare, or draft. Keep the sources, tools, and final decision visible to a person.
The Shakti case
1 buildingbecomes one bounded evidence packet4 City sourcesarrive with dates and processing receipts0 AI decisionsin address matching, counts, privacy, or the next step1 optional AI jobexplain the treated packet for human reviewThe reusable pattern
A case packet, not a data dump.
Enterprise RAG becomes useful when retrieval follows the work and its permissions. Shakti does not send a model every housing record. It builds one small packet around the building a person chose.
- 01Ask
Name the person, task, and result that matters.
- 02Retrieve
Use an approved key to fetch only relevant evidence.
- 03Treat
Normalize, redact, cap, and attach source receipts.
- 04Amplify
Let a model explain, compare, or draft within the packet.
- 05Review
Return the work and evidence to a person who decides.
The boundary is the product
AI may help with the reading. It does not own the record.
AI may not
- change source records or hide uncertainty
- expand retrieval beyond the approved case
- make eligibility, enforcement, or legal decisions
- send, publish, file, or delete without review
- turn a missing record into a confident answer
Meaningful applied use cases
The pattern travels. The boundaries change.
These are reference patterns, not finished Shakti features. Each one begins with a real task and ends with a person reviewing the work.
Prepare one repair case
Retrieve: records tied to one building ID.
Amplify: explain the timeline and draft questions with citations.
Keep human: advocacy strategy, filing, and legal judgment.
Navigate verified resources
Retrieve: current program rules and official service pages.
Amplify: compare options and explain what documents may be needed.
Keep human: eligibility decisions and applications.
Trace a policy decision
Retrieve: agendas, minutes, votes, and named budget records.
Amplify: build a dated brief and surface unanswered questions.
Keep human: interpretation, public comment, and publication.
Prepare an inspection visit
Retrieve: the assigned case, approved history, and current checklist.
Amplify: assemble the visit brief and flag conflicting evidence.
Keep human: findings, safety decisions, and enforcement.
Review evidence without losing the source
Retrieve: one application and the approved rubric.
Amplify: map submitted evidence to rubric questions.
Keep human: scoring, funding, and conflict review.
Answer policy questions with receipts
Retrieve: current, approved policies for the person's role.
Amplify: answer with exact sections and state when policies conflict.
Keep human: exceptions, discipline, and policy changes.
Local AI and traced tools
A model becomes useful when its tools are narrow and inspectable.
In the local research edition, Hermes receives the treated case packet. Its session is tagged, bounded to four turns, and separated from the public record lookup. The next engineering step is to relate each Hermes tool call to the Shakti trace without copying private content into the civic ledger.
Case packet readerSource receipt viewerTrace verifierBounded draft writer
Run the local research edition ↗
Before calling it enterprise RAG
Prove the whole loop.
- Retrieval: Can a reviewer see why every source entered the case?
- Permissions: Does the system enforce what this person and model may read?
- Treatment: Are private fields removed before model use?
- Receipts: Are source dates, tool calls, and stop reasons inspectable?
- Evaluation: Did known cases test missing, conflicting, and stale evidence?
- Review: Is the human decision explicit before anything changes outside the workspace?