EcoIQ Platform Module

EcoIQ Document Reader Agent Training Pack

Train the agent that reads bills, PDFs, reports, supplier quotes and MRV evidence.

EcoIQ Document Reader Agent Training Pack defines how the Document Reader Agent extracts reliable structured facts from uploaded documents. It trains the agent to read energy bills, fuel bills, water bills, annual reports, maintenance logs, inspection reports, technical documents, supplier quotes, invoices, MRV evidence and public disclosures without guessing missing data or creating unsupported claims.

Core purpose: Make document extraction accurate, evidence-linked and safe for Asset Passports, MRV, finance models, reports and public summaries.

Open Document Reader Training Pack Create Golden Test Case Run Extraction Evaluation Review Failed Extraction Open Required Output Schema Check PII Risk Link Document to Asset Passport Send to Human Review Train on Energy Bill Train on Supplier Quote Train on MRV Evidence

Connected EcoIQ Modules

Agent Training & Evaluation Lab

Supplies the training method, schema and evaluation pattern this pack specialises for the Document Reader Agent.

AI Agent Operations Console

Monitors Document Reader Agent task runs and confidence over time.

Data Room & Evidence Vault

Stores the source documents this agent reads and links extractions back to.

Asset Passport

Receives baseline fields extracted from bills, inspection reports and technical specs.

Impact MRV Layer

Receives baseline and after-data extracted from MRV evidence documents.

Institutional Finance Engine

Receives CAPEX/OPEX assumptions extracted from bills and supplier quotes.

Executive Briefing & Board Pack Generator

Uses extracted evidence to support investor memo and board pack claims.

Knowledge Graph & Relationship Map

Stores extracted facts as evidence nodes linked to assets and projects.

Certification & Trust Badge Engine

Uses evidence quality and missing-field detection to issue evidence badges.

Security, Privacy & Compliance Centre

Governs how detected PII and sensitive data are handled.

Public Trust & Impact Portal

Only publishes extracted facts once approved and stripped of sensitive data.

Amanah Autopilot

Flags overnight documents with weak evidence quality or missing fields for review.

Microsoft Ecosystem Core Stack

Provides the document/storage building blocks this agent can be wired into.

Why Document Reader Agent Comes First

Document Reader Agent is the foundation for EcoIQ because most downstream outputs depend on extracted document facts.

It supports

Asset Passport baseline fieldsFinance model assumptionsMRV baseline and after-dataSupplier quote comparisonInvestor memo evidenceBoard pack evidencePublic impact verificationKnowledge Graph evidence nodesTrust Badge evidence requirements

If this agent guesses or misreads data, EcoIQ can overstate savings, impact, risk or finance readiness.

Document Types

1. Energy bill

Fields to extract

SupplierBilling periodAccount/site reference if presentkWh usedUnit rateStanding chargeTotal costVAT/tax if presentMeter number if presentEstimated vs actual readingCurrencyMissing fields

2. Fuel bill

Fields to extract

Fuel typeVolume / weightUnitCostBilling periodSupplierDelivery dateAsset/siteEmissions factor required noteMissing fields

3. Water bill

Fields to extract

SupplierBilling periodVolume usedUnitTotal costMeter readingEstimated vs actualSite referenceMissing fields

4. Annual report / ESG report

Fields to extract

Company nameReporting yearEmissions dataEnergy useWater useWaste dataTargetsCapex plansClimate risksGovernance statementsSource page/sectionConfidenceMissing data

5. Maintenance log

Fields to extract

Asset nameService dateIssueAction takenDowntimeParts replacedEngineer noteRecurring issueSafety concernMissing fields

6. Inspection report

Fields to extract

SiteInspectorDateAsset conditionObserved risksRecommendationsPhotos referencedMeasurementsRequired follow-upMissing evidence

7. Supplier quote

Fields to extract

Supplier nameQuote dateTechnology/equipmentQuantityCAPEXInstallation costWarrantyLead timeExclusionsAssumptionsValidity periodPayment termsMissing fields

8. Invoice

Fields to extract

VendorInvoice dateInvoice numberItemsQuantitiesUnit pricesTotalVAT/taxCurrencyProject/site referenceMissing fields

9. Technical specification

Fields to extract

Equipment typeModelCapacityEfficiencyFuel/power requirementOperating rangeSafety notesMaintenance requirementsStandards referencedMissing fields

10. MRV evidence document

Fields to extract

ProjectBaseline periodAfter periodMeasured metricBaseline valueAfter valueUnitMethodologyEvidence qualityVerification statusMissing data

Required Output Schema

Every extraction must output:

{
  "agent_name": "Document Reader Agent",
  "document_type": "",
  "input_summary": "",
  "extracted_fields": {},
  "tables": [],
  "key_figures": [],
  "dates": [],
  "units": [],
  "currency": "",
  "asset_or_project_links": [],
  "evidence_quality": "strong | medium | weak | unreadable",
  "missing_fields": [],
  "confidence": 0.0,
  "risk_flags": [],
  "human_approval_required": true,
  "next_action": "",
  "status": "draft | needs_review | usable | blocked"
}

Rules

Never guess missing values.
Use null or "missing" where data is absent.
Preserve units exactly.
Preserve currency exactly.
Do not convert units unless explicitly requested.
If OCR/scan quality is poor, mark evidence quality as weak or unreadable.
If a number is unclear, mark it as uncertain.
If a table is present but not readable, flag it.
If the document is not related to the project, flag mismatch.
If personal data appears, flag PII risk.
If the document supports finance or MRV, require human approval.

Extraction Confidence Rules

Strong confidence

Clear machine-readable text
Complete fields
Readable tables
Clear dates and units
Document matches project/asset

Medium confidence

Most fields clear
Some missing data
Small formatting ambiguity
Document likely matches project/asset

Weak confidence

Scanned/blurred document
Unclear numbers
Missing dates/units
Document mismatch possible
Handwritten or low quality

Unreadable

Cannot reliably extract key information
Image too blurry
Table unavailable
Wrong file type
Corrupted text

Good Extraction Examples

Example 1

Input: Energy bill for Factory Line 2, March 2026.

Good output

document_type: energy bill
billing_period: March 2026
kWh_used: extracted exactly
total_cost: extracted exactly
currency: GBP
estimated_reading: yes/no if shown
missing_fields: meter number missing
evidence_quality: strong
next_action: link to Asset Passport and Finance Modelling Agent

Example 2

Input: Supplier quote for heat pump installation.

Good output

Supplier name
Quote date
Equipment
CAPEX
Installation cost
Exclusions
Validity period
Warranty
Missing fields
Human approval required before supplier recommendation

Example 3

Input: MRV after-data spreadsheet PDF.

Good output

Baseline value
After value
Unit
Method
Evidence quality
Missing reviewer approval
Status: needs_review

Bad Extraction Examples

Inventing missing kWh because total cost is known.
Changing litres to tonnes without conversion note.
Treating supplier quote as approved supplier.
Using annual report target as verified impact.
Claiming MRV Verified without after-data and human approval.

Document-Specific Safety Rules

Energy/fuel/water bills

Do not infer savings from one bill alone.
Do not claim efficiency improvement without baseline comparison.
Flag estimated meter readings.

Annual/ESG reports

Separate targets from actual results.
Separate company claims from verified data.
Flag missing methodology.

Supplier quotes

Quote is not endorsement.
Supplier due diligence is required.
Exclusions and assumptions must be extracted.

MRV evidence

Baseline and after-data both required for verification.
Estimated impact is not verified impact.
Human approval required for MRV Verified label.

Public summaries

No public claim unless approved through Public Trust Portal.
Check consent and privacy before publication.

Human Approval Triggers

Human approval required if extracted facts are used for:

Finance modelInvestor memoMRV Verified claimPublic impact storyBoard packGovernment briefSupplier recommendationIslamic finance briefMaqasid/Mizan public claimExternal report

PII and Sensitive Data Detection

Flag

NamesPhone numbersEmail addressesExact household addressSignaturesAccount numbersPersonal identifiersFaces in attached imagesPrivate financial dataConfidential supplier pricing

Output

pii_detected: true/falsepii_types: []public_safe: true/falseredaction_required: true/false

Document Reader to Knowledge Graph Mapping

DOCUMENT_EXTRACTED_FACTEVIDENCE_SUPPORTS_CLAIMDOCUMENT_LINKED_TO_ASSETDOCUMENT_LINKED_TO_PROJECTFACT_REQUIRES_REVIEWFACT_HAS_CONFIDENCEDOCUMENT_HAS_PII_RISKDOCUMENT_SUPPORTS_MRVDOCUMENT_SUPPORTS_FINANCE_MODELDOCUMENT_SUPPORTS_BADGE

Training Golden Test Cases

#	Input type	Expected fields	Expected missing	Risk flags	Human approval	Status
1	Clear UK electricity bill	Supplier, billing period, kWh, unit rate, total cost, currency	Meter number (if not shown)	None	Only if used for finance/MRV	usable
2	Poor-quality scanned fuel bill	Fuel type, supplier (partial)	Volume, cost, billing period	Low OCR quality	Yes	needs_review
3	Supplier quote with exclusions	Supplier, CAPEX, installation cost, exclusions, warranty	Payment terms (if absent)	Endorsement risk	Yes	needs_review
4	Annual report with target but no actual emissions	Company name, reporting year, target	Actual emissions data	Target vs actual confusion risk	Yes	needs_review
5	Maintenance log with recurring fault	Asset name, service date, issue, action taken	Downtime duration (if absent)	Recurring issue, possible safety concern	No	usable
6	Inspection report with missing measurements	Site, inspector, date, observed risks	Measurements	Missing evidence	No	needs_review
7	MRV baseline but no after-data	Project, baseline period, baseline value, unit	After period, after value	Cannot verify impact yet	Yes	blocked
8	Invoice with VAT and multiple line items	Vendor, invoice number, items, VAT, total	Project/site reference (if absent)	None	No	usable
9	Technical spec with capacity and efficiency	Equipment type, model, capacity, efficiency	Standards referenced (if absent)	None	No	usable
10	Wrong document uploaded to project	Document type only	All project-specific fields	Document/project mismatch	Yes	blocked

Evaluation Metrics

Field extraction accuracyMissing field detectionUnit preservation accuracyDocument type classification accuracyTable extraction qualityPII detection rateUnsupported claim rateHuman approval trigger accuracyJSON/schema validityReviewer acceptance rate

Prompt Template

System prompt

You are the EcoIQ Document Reader Agent. Extract only facts that are visible in the provided document. Do not guess missing values. Preserve units, dates and currency. Label uncertainty clearly. Separate targets from actual results. Separate estimates from verified data. Flag PII and sensitive information. Require human approval when extracted data supports finance, MRV, public reporting, supplier recommendation or high-impact decisions.

Task prompt

Read this document and return structured extraction using the required schema. Identify document type, extracted fields, missing fields, evidence quality, PII risk, confidence, human approval requirement and next action.

No Harm Gate for Document Reading

Before using extraction downstream, check:

Is the document type correctly identified?
Are units preserved?
Are missing fields clearly marked?
Are unclear numbers flagged?
Is PII detected?
Is the document linked to the correct asset/project?
Is evidence quality strong enough?
Are targets separated from actuals?
Is estimated data separated from verified data?
Is human approval required before downstream use?

Safety and Governance

Document Reader Agent extracts facts; it does not verify truth by itself.
Missing data must not be guessed.
Extracted values need human review when used for finance, MRV, public reporting or high-impact recommendations.
Supplier quotes are not supplier endorsements.
Annual report targets are not verified impact.
Maqasid/Mizan is ethical decision-support, not a fatwa.
Sensitive and personal data must be protected.