Direct answer for “why natural-language spatial queries fail”: AI does not replace ERP, EAM or the asset register. Its practical value is to connect physical observations from RFID, inspection or GIS to trusted records, then use RAG and an LLM to explain an exception with citations. Permission checks, deterministic validation and human approval still decide what may be written back.

What problem does this solve?

Municipal teams need to answer why natural-language spatial queries fail, but map layers, permits, work orders and inspection records have different owners, dates, coordinate systems and access rules.

Operational truth is usually distributed across EAM, ERP, DMS, email, inspection forms, RFID events and GIS. Without a common identifier, timestamp and owner, a team cannot reliably distinguish a missing asset from a moved asset or an asset that simply produced no trusted read.

Adding an LLM to uncontrolled data can increase risk. It may produce a fluent explanation from an old record or an incorrectly permitted source. The programme should therefore start with identity, event quality and governance—not language generation.

How the solution works

Use a governed spatial evidence pipeline with coordinate validation, metadata and permission filtering, GIS query execution, Hybrid Search, Reranking and LLM explanation before any workflow action.

Build a connector and ingestion layer through an API Gateway or secure adapters. Apply OCR to Arabic forms where required, and extract metadata such as location, date, document type, confidence and owning department.

Carry permissions with each record or chunk before Chunking and Embeddings. Use Hybrid Search combining BM25 and Semantic Search, followed by Reranking, so the LLM receives only permitted and relevant context.

Let the LLM interpret the question, assemble evidence and explain the exception or next step. It must not decide a legal or financial asset state, or claim an inventory result that the source events do not prove.

Send the recommendation to a deterministic rule engine for RBAC or ABAC, duplicate checks, timestamp checks and workflow validation, then to a human approver before any EAM or ERP write-back.

Step by step

  1. Capture an observation Capture an RFID event, inspection result or GIS update with identifier, timestamp and event source.
  2. Resolve identity Match the identifier to EAM or ERP and make conflicts visible instead of hiding them.
  3. Constrain context Apply IAM and RBAC/ABAC before indexing the permitted text and metadata.
  4. Retrieve evidence Use Hybrid Search and Reranking to find records the user may access that answer the question.
  5. Explain and route The LLM produces a cited explanation, then the recommendation passes to rules and human approval.

Key capabilities

  • Exception discovery Identify assets where a read, location or work order conflicts with the accepted record.
  • Grounded enterprise search Find relevant documents and observations with citation accuracy instead of an unsupported answer.
  • Arabic and English explanation Explain the case to operators while keeping the LLM in a context and reasoning role, not as the system of record.
  • Decision routing Turn a recommendation into a review task for the authorised owner with a clear audit trail.

Integrations

  • EAM / ERP Exchange asset ID, state, cost centre, work orders and transfers; write changes only after validation and approval.
  • RFID / RTLS Ingest reader, antenna, timestamp and confidence events with duplicate suppression and read-zone rules.
  • GIS Link the asset to a facility, zone or geographic location and show the last trusted observation.
  • DMS and Arabic OCR Parse Arabic forms and reports, bind them to the asset and apply classification and restricted access.
  • IAM / SSO Enforce RBAC and ABAC, recording the user, tool and reason for each access or approval.

Industry use cases

  • Government facilities Confirm the location of maintenance equipment, pumps and generators before dispatching a field team.
  • Warehouses and procurement Match received and transferred assets with purchase, transfer and custody records.
  • Inspection and compliance Assemble observation, document and approval evidence into a reviewable audit response.

Implementation approach

  1. Choose a pilot scope Select one asset class and site, and define the decision you want to accelerate or make provable.
  2. Clean identity Standardise asset ID, location, owner and state; record conflicts instead of silently deleting them.
  3. Measure observation quality Test RFID, inspection or GIS coverage, duplication, timestamps and confidence before building RAG.
  4. Build governed retrieval Apply metadata and permission filtering, Hybrid Search and Reranking against a set of real questions.
  5. Define the decision boundary Separate the LLM recommendation from rule validation, human approval and any EAM or ERP write-back.
  6. Operate and observe Track Groundedness, Citation Accuracy, response time, escalation rate and approved changes.

Limitations & prerequisites

  • An LLM cannot prove physical presence without a trusted observation.
  • RFID is affected by metal, environment, antennas and read zones, so it needs RF engineering and acceptance tests.
  • Arabic OCR can misread asset IDs and dates; use confidence thresholds and human review for critical fields.
  • RAG cannot repair contradictory records or duplicate identity; it should expose the conflict and request a data-owner decision.

FAQ

  • Does an LLM replace EAM in asset inventory? No. EAM remains authoritative for accepted transactions and state. The LLM searches, explains and assembles context; rules and approval decide what is written.
  • Is RFID alone enough for continuous inventory? No. RFID is an observation source that requires RF engineering, suitable tags, event processing, common asset identity and an operational system connection.
  • How do we prevent permission leakage into RAG? Apply IAM and RBAC or ABAC before indexing and retrieval, carry permission metadata with each chunk and test negative access cases before go-live.
  • When should a Private LLM or On-Prem AI be used? When record sensitivity, Data Residency or policy requires data and inference to remain in a controlled environment. The decision should follow classification and controls, not fashion.
  • Which KPIs show that the system works? Measure answer time, Citation Accuracy, Groundedness, correct exception escalation, case closure time and the share of changes passing rules and approval.
  • Can an agent execute an asset transfer automatically? It can propose Tool Calling or create a task, but transfer, disposal or financial changes need deterministic validation, permission, approval and an audit trail under policy.