features_details

Features

Feature 1: Charter Validation

Validates contributions against the charter rules.

Property	Value
Class	CharterValidationFeature
Prompt	forseti.charter_validation
Opik Name	forseti461-user-charter-validation
Variables	title, body
Output	ValidationResult
Registered	Auto (always)

Checks for violations:

• Personal attacks or discrimination
• Spam or advertising
• Off-topic content (unrelated to Audierne-Esquibien)
• False information

Identifies encouraged aspects:

• Concrete proposals
• Constructive criticism
• Questions and clarifications
• Shared expertise

Feature 2: Category Classification

Assigns contributions to one of 7 predefined categories.

Property	Value
Class	CategoryClassificationFeature
Prompt	forseti.category_classification
Opik Name	forseti461-user-category-classification
Variables	title, body, current_category_line
Output	ClassificationResult
Registered	Auto (always)

Categories:

• economie - Commerce, tourism, port, fishing
• logement - Housing, urban planning
• culture - Heritage, events, associations
• ecologie - Environment, energy, biodiversity
• associations - Local organizations
• jeunesse - Youth, schools, education
• alimentation-bien-etre-soins - Food, health, wellness

Feature 3: Wording Correction

Suggests improvements for clarity and constructiveness.

Property	Value
Class	WordingCorrectionFeature
Prompt	forseti.wording_correction
Opik Name	forseti461-user-wording-correction
Variables	title, body
Output	WordingResult
Registered	Optional (enable_wording=True)

Improvements include:

• Clarity and readability
• Grammar corrections
• More constructive phrasing
• Removing inflammatory language

Feature 4: Anonymization

PII (Personally Identifiable Information) protection with three complementary modes.

Property	Value
Class	AnonymizationFeature
Prompt	forseti.anonymization
Variables	text
Output	AnonymizationResult
Registered	Manual (standalone execution)

Three modes (the Anonymization Trilemma):

Transcript Mode (Deterministic, Regex)

For Plaud AI and timestamped meeting transcripts. Free, instant, deterministic:

00:00:00 Florent Lardic    →    00:00:00 Speaker_1
00:05:23 Malika Redaouia   →    00:05:23 Speaker_2

• Consistent Speaker_N assignment across the document
• Fuzzy matching for spelling variations (Levenshtein distance)
• Inline mention replacement
• Cost: Free | Speed: Instant | Accuracy: High for structured formats

LLM Mode (Experimental)

For general documents requiring context understanding:

• Names → [PERSONNE_N], emails → [EMAIL_N], phones → [TELEPHONE_N], addresses → [ADRESSE_N]
• Extracts non-PII keywords (organizations, places) for theme analysis
• Cost: ~$0.001-0.003 per doc | Speed: 2-5s | Accuracy: High

Opik PII Guardrail as alternative: opik.guardrails.guards.pii.PII provides NLP-based entity detection (50-100ms, free) that may be a more efficient replacement for LLM mode in many cases. Currently used for post-processing validation, but could serve as primary anonymizer for simpler documents. See the Anonymization Trilemma blog post for cost/accuracy/speed tradeoffs.

Auto Mode (Default)

Detects document type and routes automatically:

• Transcript with speaker names → Regex mode
• Already anonymous → Skip
• General document → LLM mode
• All modes → PII validation (Opik Guardrail)

Entity types anonymized:

Entity	Placeholder	Example
Person	[PERSONNE_N]	Jean Dupont → [PERSONNE_1]
Email	[EMAIL_N]	[email protected] → [EMAIL_1]
Phone	[TELEPHONE_N]	06 12 34 56 78 → [TELEPHONE_1]
Address	[ADRESSE_N]	12 rue de la Paix → [ADRESSE_1]

Preserved as keywords (not anonymized): organizations, public places, institutions.

Files:

File	Description
app/agents/forseti/features/anonymization.py	LLM-based feature
app/mockup/anonymizer.py	Transcript regex, type detection, PII validation
app/mockup/field_input.py	Pipeline integration

Feature 5: Translation (Available, Not Integrated)

Translates French contributions to English.

Property	Value
Class	TranslationFeature
Variables	French constat + idees
Output	TranslationResult
Registered	Not integrated
Status	Defined, no active usage

Batch Validation (Experiments)

Batch validation is handled as Opik experiments, not as a feature class. This allows:

• A/B testing different prompt versions
• Tracking metrics across validation runs
• Comparing performance over time

See Prompt Management for experiment setup.

Feature Summary

#	Feature	Registration	Status
1	Charter Validation	Auto	Active
2	Category Classification	Auto	Active
3	Wording Correction	Optional	Active if enabled
4	Anonymization	Manual	Active (standalone)
5	Translation	Not integrated	Available