AI Integrated Chatbot Portal

AI Integrated Chatbot Portal is a secure PHP 8.3 and MySQL 8 platform for institutional chatbot deployments. It provides a multi-provider LLM gateway, retrieval-augmented generation, prompt versioning, dynamic branding, analytics, and an administration layer designed for controlled rollout across departments.

Key Capabilities

Area	Capability
Multi-LLM gateway	Route requests to OpenAI, Gemini, or DeepSeek with provider health checks, fallback, timeout budgets, and per-group defaults
Intent-aware routing	Classify user intent and adapt provider order, sampling settings, RAG use, and output budget per request
Prompt firewall	Detect prompt extraction, secret requests, data exfiltration attempts, and unsafe administrative actions before provider calls
Chat experience	Modern operations-console UI with compact navigation, quick prompts, live safety/routing signals, usage telemetry, and source citations
RAG knowledge base	Upload TXT, PDF, and DOCX sources, chunk content, create embeddings, store vectors in MySQL, retrieve cited context, and track document provenance
Knowledge-base source license audit	Executable review of source reuse rights, citation approval, redistribution, provider-training limits, personal-data approval, access scope, and review freshness
RAG freshness audit	Review source ownership, review dates, citation requirements, and stale indexes before release
Knowledge-base access scope audit	Gate RAG collections by bot, department, classification, retrieval scope, citation requirement, personal-data approval, and review date
RAG citation integrity audit	Check whether answer claims cite retrieved, approved, current, high-score, and citation-allowed sources
Answer provenance drift audit	Compare current answer citations against a baseline to detect missing expected sources, stale reviews, retrieval-score drops, and citation-set drift
RAG source retirement	Runbook for retiring, replacing, restricting, re-owning, or quarantining knowledge-base sources while preserving citation and audit history
Prompt operations	Versioned system prompts, personas, release notes, approval workflow, rollback, and A/B experiment metadata
Prompt release audit	Gates prompt and persona releases for approval, evaluation, red-team, rollback, RAG freshness, tool-policy, and human-review evidence
Evaluation lab	JSON evaluation packs and CLI runner for prompt, RAG, refusal, citation, and policy behavior checks
Evaluation coverage gate	Audits scenario packs for required safety, citation, policy, credential, and research coverage before release
Prompt injection regression audit	Checks release-blocking prompt-injection packs for direct, indirect, extraction, RAG poisoning, tool abuse, and credential-safety coverage
Human review queue audit	Release and closure gate for high-risk chatbot conversations, reviewer assignment, SLA, redaction, escalation, evidence hashes, and version context
Administration	RBAC for super admins, department admins, reviewers, and end users with scoped permissions
Admin activity evidence audit	Governance gate for exported prompt, provider, RAG, RBAC, branding, cost, retention, and decommissioning changes
Tool invocation policy audit	Release gate for chatbot tool permissions, high-impact actions, human approval, logging, credential scope, rollback, and emergency stop readiness
Branding	Dashboard-managed logo, palette, typography, support links, and dynamic CSS variables without code edits
Security	MFA-ready admin accounts, encrypted provider credentials, CSRF protection, rate limiting, audit logging, retention controls, and secure headers
Conversation retention policy audit	Executable review of prompt, output, RAG, embedding, tool-call, audit-export, redaction, deletion, legal-hold, and provider-training retention controls
Prompt log minimization	Pre-persistence redaction for emails, identifiers, tokens, API keys, phone numbers, and long secret-like values
Audit export	Redacted evidence packages with package-level SHA-256 integrity checks for conversation review
Redaction assurance	Residual scan for unredacted tokens, API keys, JWT-like values, emails, and direct identifiers before evidence sharing
Provider incident evidence	Redacted, hashable packages for provider degradation, fallback routing, cost spikes, and safety-filter changes
Provider failover readiness	Pre-release audit for fallback provider order, health checks, timeout/retry limits, residency and safety equivalence, logging, runbook, and evidence
Provider model deprecation readiness	Executable lifecycle audit for retiring models, replacement evidence, fallback testing, prompt and RAG compatibility, safety equivalence, cost impact, and user communication
Model migration canary readiness	Executable rollout audit for limited traffic, shadow evaluation, locked baselines, rollback thresholds, safety and citation regressions, cost deltas, and monitoring ownership
Provider data processing review	Review provider data categories, training use, retention, subprocessors, regions, deletion, assurance evidence, fallback eligibility, and portal configuration controls
Cost budget audit	Department and bot-level cost guardrail for projected spend, hard stops, owner review, and approval evidence
Bot decommissioning	Runbook for retiring chatbot instances while preserving required evidence, removing access, handling RAG sources, revoking credentials, and closing audit records
Governance	Interaction logging, cost tracking, token analytics, provider uptime, moderation flags, data residency controls, and configurable retention windows
Operations	Docker Compose stack, health endpoint, CI syntax validation, migration SQL, and deployment hardening guidance

Architecture

flowchart LR
    Browser["Chat UI and Admin UI"] --> Router["PHP Front Controller"]
    Router --> Auth["Auth, RBAC, MFA"]
    Router --> Chat["Chat Orchestrator"]
    Chat --> Prompt["Prompt Version Manager"]
    Chat --> Rag["RAG Retriever"]
    Rag --> Vector["MySQL Vector Store"]
    Chat --> Gateway["LLM Provider Router"]
    Gateway --> OpenAI["OpenAI"]
    Gateway --> Gemini["Gemini"]
    Gateway --> DeepSeek["DeepSeek"]
    Chat --> Logs["Usage, Cost, Audit Logs"]
    Admin["Admin Services"] --> Branding["Branding Engine"]
    Admin --> Settings["Provider and Bot Settings"]

Loading

Repository Structure

Path	Purpose
public/	Front controller and frontend assets
src/AI/	Provider abstraction, fallback routing, and chat orchestration
src/Rag/	Document extraction, chunking, embeddings, vector retrieval, and citations
src/Security/	RBAC, MFA, CSRF, rate limiting, encryption, and audit helpers
src/Admin/	Branding, prompt, provider, and configuration services
src/Analytics/	Token, cost, latency, and uptime recording
database/schema.sql	MySQL 8 schema for users, chats, RAG, prompts, analytics, and audit logs
docs/	Architecture, security, API, RAG, and deployment guidance

Quick Start

cp .env.example .env
docker compose up --build

Open the portal at http://localhost:8080.

Run local checks:

php scripts/lint.php
php scripts/run-evaluation.php
php scripts/evaluation-coverage-gate.php
php scripts/prompt-injection-regression-audit.php
php scripts/answer-provenance-drift-audit.php
php scripts/kb-source-license-audit.php
php scripts/rag-freshness-audit.php
php scripts/check-security-headers.php
php scripts/provider-incident-evidence.php
php scripts/provider-failover-readiness.php
php scripts/provider-model-deprecation-readiness.php
php scripts/model-migration-canary-readiness.php
php scripts/prompt-release-audit.php
php scripts/human-review-queue-audit.php
php scripts/admin-activity-evidence-audit.php
php scripts/tool-invocation-policy-audit.php
php scripts/conversation-retention-policy-audit.php
php scripts/cost-budget-audit.php
php scripts/prompt-log-redaction-preview.php
php scripts/redaction-residual-audit.php
php scripts/kb-access-scope-audit.php
php scripts/rag-citation-integrity-audit.php

Configuration Highlights

APP_ENV=local
APP_KEY=base64:replace-with-32-byte-sodium-key
DB_HOST=mysql
DB_DATABASE=ai_chatbot_portal
OPENAI_API_KEY=
GEMINI_API_KEY=
DEEPSEEK_API_KEY=
DEFAULT_PROVIDER=openai
DATA_RESIDENCY_REGION=UAE
AUDIT_RETENTION_DAYS=365

Provider keys are encrypted before storage. Environment variables may be used for local development, but production should use the encrypted provider_credentials table or a managed secret store.

Security Model

• Admin routes require authenticated sessions, CSRF validation, MFA enrollment, and RBAC checks.
• Provider credentials are encrypted with sodium secretbox using APP_KEY.
• All administrative changes are written to append-only audit logs.
• Chat requests are rate-limited by user, IP, provider, and bot instance.
• Prompt versions are immutable once approved, enabling rollback and post-incident review.
• Knowledge-base documents keep provenance metadata so generated answers can cite source chunks.
• Conversation, prompt, output, RAG, embedding, tool-call, and audit-export stores should pass the retention policy audit before production logging is enabled.

Core Workflows

1. Super admin configures provider credentials and fallback order.
2. Department admin creates a bot instance with persona, prompt version, model settings, and branding.
3. Knowledge manager uploads documents for indexing.
4. End users chat with the bot through a branded interface.
5. The orchestrator retrieves relevant context, routes the request to the selected provider, records usage, and returns cited answers.
6. Analytics dashboards show token spend, provider uptime, latency, flagged conversations, and RAG hit rates.

Documentation

• Architecture
• Security Model
• RAG Pipeline
• Knowledge-Base Source License Audit
• RAG Freshness Audit
• Knowledge-Base Access Scope Audit
• RAG Citation Integrity Audit
• Answer Provenance Drift Audit
• RAG Source Retirement Runbook
• Innovation Layer
• Evaluation Lab
• Evaluation Coverage Gate
• Prompt Injection Regression Audit
• Human Review Queue Audit
• Admin Activity Evidence Audit
• Tool Invocation Policy Audit
• Conversation Retention Policy Audit
• Conversation Audit Export
• Redaction Residual Audit
• Provider Incident Evidence
• Provider Failover Readiness
• Provider Model Deprecation Readiness
• Model Migration Canary Readiness
• Provider Data Processing Review
• Prompt Release Audit
• Bot Decommissioning Runbook
• Cost Budget Audit
• Prompt Log Redaction
• API Reference
• Deployment Guide
• Operations Runbook

Roadmap

• SSO integration with SAML/OIDC.
• Runtime human-review queues for flagged conversations.
• Department-level cost budget enforcement in runtime routing.
• Evaluation harness for prompt and model regression testing.
• Live SSE streaming with provider-specific adapters.
• Optional managed vector database adapters.

License

MIT License. See LICENSE.