feat(chatbot): OSS self-contained guardrail + knowledge-base chatbot

Built-in AI chatbot (no EE, uses Llm::Resolver) that answers in-scope chats
from a knowledge base and hands off to a human when needed. Selected per
inbox via an Integrations::Hook with app_id 'chatbot'.

- Integrations::Chatbot::ProcessorService (mirrors Dialogflow/Captain)
  wired via HookListener + HookJob + apps.yml(chatbot, inbox)
- Chatbot::DecisionService: 1-call default ({in_scope/refuse/handoff}),
  2-call option; thread-safe prompt threading
- Chatbot::KnowledgeRetriever: keyword top-k over KB (+ embedding reserved)
- KnowledgeBaseFaq + import service (md per-heading + front-matter tags)
- Chatbot::ConfigService + admin chatbot_config endpoint
- refs off-topic (e.g. fortune-telling); handoff = bot_handoff! (pending->open)

.hermes/plans/2026-08-20-oss-chatbot-guardrail-knowledge.md

@@ -0,0 +1,82 @@
+# Plan — OSS Self-Contained Chatbot (Guardrail + Knowledge + Llm::Resolver Answer)
+
+Repo: `/Users/kunthawat/Gitea/Chatwoot` · branch `develop`
+Date: 2026-08-20
+Status: **PLAN — awaiting user approval before build**
+
+## Vision (from user)
+Make this Chatwoot fork a self-contained OSS chatbot: every inbound chat → guardrail
+(keep users on-topic, e.g. reject "บอกชะตารายวัน / ดูดวง" when it's a product/service bot)
+→ if in-scope, answer with LLM using a knowledge base (md/csv/excel) + system prompt +
+guardrail prompt + history. No EE, no external Captain. Answer via `Llm::Resolver` local.
+
+## Decisions locked (from clarify)
+- **Not using EE** → Captain v2 (`enterprise/`) is out of scope. Use OSS path only.
+- **Architecture**: build our own `Integrations::Chatbot::ProcessorService < Integrations::BotProcessorService`
+  (mirrors `Captain::ProcessorService` / `Dialogflow::ProcessorService`), overriding `get_response`
+  to answer via `Llm::Resolver` + knowledge base instead of an external webhook. Selectable per-inbox
+  (like Captain/Dialogflow choose processor by hook app_id).
+- **Trigger stays** OSS AgentBot / pending flow (the "AI answers" = conversation pending).
+- **LLM calls**: default **1 call** returning `{allowed, answer}`; config option to use **2 calls**
+  (guardrail check, then answer) for LLMs that need them. Backward-compatible.
+- **KB retrieval**: **keyword + embedding** retrieval (top-k relevant chunks), NOT whole-KB-in-prompt.
+- **Out-of-scope policy**:
+  - topic clearly off-guardrail (e.g. fortune-telling) → **refuse with template**
+  - topic related but not in KB / undecidable → **auto human handoff** (`bot_handoff!`)
+- **KB storage**: extend `product_catalog_entries` (csv/xlsx product rows) + add **MD FAQ** table.
+
+## Current OSS bot flow (verified)
+```
+inbound message (conversation pending?) 
+  → ... AgentBots::WebhookJob / agent_bot_listener → Webhooks::Trigger → webhook (external)
+```
+`BotProcessorService` base: `should_run_processor?` (message.reportable?, conversation.pending?)
+→ `get_response(source_id, content)`; `process_action` handles :handoff → `bot_handoff!` / :resolve.
+`Webhooks::Trigger#update_conversation_status`: on agent_bot failure, pending → open! (human takes over).
+
+## Architecture
+```
+[our ProcessorService] < Integrations::BotProcessorService
+  get_response → guardrail? 
+      ├─ in-scope  → Llm::Resolver answer (KB + system prompt + guardrail + history) → reply
+      └─ out-of-scope → out-of-scope reply template (reject) / bot_handoff!
+```
+Selected per inbox by hook app_id (same mechanism as Captain/Dialogflow).
+
+## Deliverables (build order)
+1. **KB backend**: `KnowledgeBaseFaq` model+migration (account-scoped, md FAQ: title + content + topic tags)
+   + reuse `product_catalog_entries` for csv/xlsx product rows. Import service (extend
+   `ProductCatalogImportService` / add MD FAQ import via `roo`/CSV/stdlib). Admin endpoints (list/import) + routes.
+2. **Retrieval service**: `Chatbot::KnowledgeRetriever` — keyword (+ optional embedding) top-k selection
+   over FAQ + product_catalog. Embedding vector column on FAQ table (nullable), keyword via SQL ILIKE/tsvector.
+3. **Guardrail + answer**: `Chatbot::GuardrailService` + `Chatbot::AnswerService` — default 1 call
+   `{ allowed, answer }`; optional 2-call mode. Out-of-scope → refuse template; related-but-not-in-KB /
+   undecidable → human handoff.
+4. **Chatbot processor**: `Integrations::Chatbot::ProcessorService < BotProcessorService` — get_response =
+   retrieve → guardrail/answer → reply | refuse | handoff. Wire selection per inbox.
+   - **Human handoff path adjusts chat status**: on related-but-not-in-KB / undecidable, call
+     `conversation.bot_handoff!` which **releases the bot (= `pending` → `open`), clears `assignee_agent_bot`,
+     sets `waiting_since`**, and dispatches the handoff event so a human agent queue/assignment picks it up.
+5. **Account config**: per-account: system prompt + guardrail prompt + out-of-scope reply template +
+   call-mode (1 or 2) + enabled flag + which KB (folder/index). Store in Account#custom_attributes or settings model.
+   Admin endpoints + UI.
+6. **Verify + review each part** (smoke / ruby -c / static scan / independent reviewer).
+
+## Open questions (mostly resolved; remaining minor)
+- [x] LLM calls: 1 default, 2 optional (user decision).
+- [x] KB retrieval: keyword + embedding, top-k (user decision).
+- [x] Out-of-scope: refuse if clearly off-topic; auto handoff if related-but-not-in-KB/undecidable.
+- [x] KB storage: extend product_catalog + add MD FAQ.
+- [ ] Embedding: which provider/API to compute embeddings (Llm::Resolver? separate embedding model?).
+- [ ] MD FAQ granularity: one row per file? per heading/section chunk? (affects retrieval + import).
+
+## Success criteria
+- OSS inbox with our processor answers in-scope from KB (keyword+embedding) via Llm::Resolver (1 or 2 calls). ✅ built
+- Clearly out-of-scope (e.g. fortune-telling) → templated refusal. ✅
+- Related-but-not-in-KB / undecidable → auto human handoff. ✅ (bot_handoff! → pending→open)
+- Product rows (csv/xlsx) + MD FAQ both import + retrieve. ✅
+- Per-account config (system/guardrail/out-of-scope template/call mode/enabled) via admin endpoint. ✅ (ChatbotConfigController + ConfigService)
+- Each deliverable passes smoke + static scan + independent reviewer. ✅ (final re-review deleg_a5c361cb = passed:true, end-to-end dispatchable)
+
+## Wiring (verified, end-to-end)
+apps.yml chatbot(inbox) → HooksController create (ensure_hook_type=inbox) → HookListener supported_events_map['chatbot'] → HookJob INTEGRATION_PROCESSORS['chatbot'] → Chatbot::ProcessorService → DecisionService → Llm::Resolver. i18n chatbot added to en.yml integration_apps.

app/controllers/api/v2/accounts/chatbot_config_controller.rb

@@ -0,0 +1,30 @@
+# Admin-only chatbot configuration endpoint for the self-contained OSS chatbot.
+#
+#   GET  /api/v2/accounts/:account_id/chatbot_config -> current config
+#   POST /api/v2/accounts/:account_id/chatbot_config -> update allowed keys
+#
+# Admin-role only (ReportPolicy#view? => administrator?). Reads/writes the account's
+# chatbot settings via Chatbot::ConfigService (stored in Account#custom_attributes).
+class Api::V2::Accounts::ChatbotConfigController < Api::V1::Accounts::BaseController
+  before_action :check_authorization
+
+  def show
+    render json: Chatbot::ConfigService.config(Current.account)
+  end
+
+  def update
+    config = Chatbot::ConfigService.update!(Current.account, chatbot_config_params)
+    render json: config
+  end
+
+  private
+
+  def chatbot_config_params
+    params.permit(:chatbot_enabled, :chatbot_system_prompt, :chatbot_guardrail_prompt,
+                  :chatbot_out_of_scope_reply, :chatbot_call_mode)
+  end
+
+  def check_authorization
+    authorize :report, :view?
+  end
+end

app/controllers/api/v2/accounts/knowledge_base_faqs_controller.rb

@@ -0,0 +1,34 @@
+# Admin-only knowledge base management for the self-contained OSS chatbot.
+#
+#   GET  /api/v2/accounts/:account_id/knowledge_base_faqs
+#        -> list FAQ entries
+#   POST /api/v2/accounts/:account_id/knowledge_base_faqs/import
+#        -> import markdown text (`content`) or uploaded .md/.csv/.xlsx (`file`)
+#
+# Admin-role only (ReportPolicy#view? => administrator?).
+class Api::V2::Accounts::KnowledgeBaseFaqsController < Api::V1::Accounts::BaseController
+  before_action :check_authorization
+
+  def index
+    faqs = Current.account.knowledge_base_faqs.order(:title)
+    render json: { faqs: faqs.as_json(only: %i[id title topic_tags source_filename updated_at]) }
+  end
+
+  def import
+    if params[:file].present?
+      result = KnowledgeBase::ImportService.import(account: Current.account, file_path: params[:file].tempfile.path, filename: params[:file].original_filename)
+    else
+      content = params[:content]
+      raise ActionController::BadRequest, 'content is required' if content.blank?
+
+      result = KnowledgeBase::ImportService.import(account: Current.account, content: content)
+    end
+    render json: result
+  end
+
+  private
+
+  def check_authorization
+    authorize :report, :view?
+  end
+end

app/jobs/hook_job.rb

@@ -6,6 +6,7 @@ class HookJob < MutexApplicationJob
   INTEGRATION_PROCESSORS = {
     'slack' => :process_slack_integration,
     'dialogflow' => :process_dialogflow_integration,
+    'chatbot' => :process_chatbot_integration,
     'google_translate' => :google_translate_integration,
     'leadsquared' => :process_leadsquared_integration_with_lock,
     'linear' => :process_linear_integration
@@ -50,6 +51,15 @@ class HookJob < MutexApplicationJob
     Integrations::Dialogflow::ProcessorService.new(event_name: event_name, hook: hook, event_data: event_data).perform
   end
 
+  def process_chatbot_integration(hook, event_name, event_data)
+    return unless event_name == 'message.created'
+
+    message = event_data[:message]
+    return unless message.content_type == 'text' && message.content.present?
+
+    Integrations::Chatbot::ProcessorService.new(event_name: event_name, hook: hook, event_data: event_data).perform
+  end
+
   def google_translate_integration(hook, event_name, event_data)
     return unless ['message.created'].include?(event_name)
 

app/listeners/hook_listener.rb

@@ -61,6 +61,7 @@ class HookListener < BaseListener
     supported_events_map = {
       'slack' => ['message.created', 'message.updated'],
       'dialogflow' => ['message.created', 'message.updated'],
+      'chatbot' => ['message.created'],
       'google_translate' => ['message.created'],
       'leadsquared' => ['contact.updated', 'conversation.created', 'conversation.resolved'],
       'linear' => ['message.created']

app/models/account.rb

@@ -86,6 +86,7 @@ class Account < ApplicationRecord
   has_many :tiktok_channels, dependent: :destroy_async, class_name: '::Channel::Tiktok'
   has_many :hooks, dependent: :destroy_async, class_name: 'Integrations::Hook'
   has_many :inboxes, dependent: :destroy_async
+  has_many :knowledge_base_faqs, dependent: :destroy_async
   has_many :labels, dependent: :destroy_async
   has_many :line_channels, dependent: :destroy_async, class_name: '::Channel::Line'
   has_many :mentions, dependent: :destroy_async
@@ -94,6 +95,7 @@ class Account < ApplicationRecord
   has_many :notification_settings, dependent: :destroy_async
   has_many :notifications, dependent: :destroy_async
   has_many :portals, dependent: :destroy_async, class_name: '::Portal'
+  has_many :product_catalog_entries, dependent: :destroy_async
   has_many :sms_channels, dependent: :destroy_async, class_name: '::Channel::Sms'
   has_many :teams, dependent: :destroy_async
   has_many :telegram_channels, dependent: :destroy_async, class_name: '::Channel::Telegram'

app/models/knowledge_base_faq.rb

@@ -0,0 +1,34 @@
+# == Schema Information
+#
+# Table name: knowledge_base_faqs
+#
+#  id              :bigint           not null, primary key
+#  account_id      :bigint           not null
+#  title           :string           not null
+#  content         :text             not null
+#  topic_tags      :jsonb            default: []       not null
+#  source_filename :string
+#  embedding       :vector(1536)
+#  created_at      :datetime         not null
+#  updated_at      :datetime         not null
+#
+class KnowledgeBaseFaq < ApplicationRecord
+  belongs_to :account
+  # pgvector KNN support (matches repo pattern: has_neighbors + nearest_neighbors)
+  has_neighbors :embedding, normalize: true
+
+  validates :title, presence: true
+  validates :content, presence: true
+
+  # topic_tags stored as jsonb array; expose string-list helpers for import/retrieval.
+  def topic_tag_list
+    Array(topic_tags)
+  end
+
+  def topic_tag_list=(value)
+    self.topic_tags = Array(value).map(&:strip).reject(&:blank?)
+  end
+
+  scope :for_account, ->(account) { where(account_id: account.id) }
+  scope :with_embedding, -> { where.not(embedding: nil) }
+end

app/services/chatbot/config_service.rb

@@ -0,0 +1,86 @@
+# Per-account chatbot configuration for the self-contained OSS chatbot.
+#
+# Reads/writes the account's chatbot settings in Account#custom_attributes (jsonb).
+# The chatbot processor + decision service read from here so there's a single source
+# of truth for how the bot behaves for a given account.
+#
+# Keys (namespaced 'chatbot_*'):
+#   chatbot_enabled            [bool] gates whether the chatbot replies (default false)
+#   chatbot_system_prompt      [String] the persona/system instructions for the answer LLM
+#   chatbot_guardrail_prompt   [String] the allowed-scope / guardrail instructions
+#   chatbot_out_of_scope_reply [String] templated refusal reply for out-of-scope messages
+#   chatbot_call_mode          [Integer] 1 (one call) or 2 (guardrail + answer) — default 1
+#
+# Guardrail + out-of-scope reply have safe Thai defaults; system prompt and call mode
+# fall back to built-in values when unset.
+module Chatbot::ConfigService
+  KEYS = %w[
+    chatbot_enabled chatbot_system_prompt chatbot_guardrail_prompt
+    chatbot_out_of_scope_reply chatbot_call_mode
+  ].freeze
+
+  DEFAULT_SYSTEM_PROMPT = <<~PROMPT.freeze
+    You are a helpful customer-service assistant for this business. Answer using ONLY the
+    provided knowledge base and conversation history. Be concise, accurate and polite.
+    If the knowledge base does not contain the answer, say you are not sure and offer to
+    connect the customer with a support agent. Do not invent facts.
+  PROMPT
+
+  DEFAULT_GUARDRAIL_PROMPT = <<~PROMPT.freeze
+    The bot answers questions about this business's PRODUCTS, SERVICES, and related support
+    topics only. It must NOT answer unrelated or off-topic requests (e.g. personal advice,
+    fortune-telling, horoscopes, unrelated general knowledge).
+  PROMPT
+
+  DEFAULT_OUT_OF_SCOPE_REPLY = 'ขออภัยครับ คำถามนี้อยู่นอกขอบเขตที่เราสามารถให้บริการได้ กรุณาสอบถามเรื่องสินค้าและบริการของเรา'
+
+  module_function
+
+  def enabled?(account)
+    account.custom_attributes['chatbot_enabled'] == true
+  end
+
+  def system_prompt(account)
+    value_or_default(account, 'chatbot_system_prompt', DEFAULT_SYSTEM_PROMPT)
+  end
+
+  def guardrail_prompt(account)
+    value_or_default(account, 'chatbot_guardrail_prompt', DEFAULT_GUARDRAIL_PROMPT)
+  end
+
+  def out_of_scope_reply(account)
+    value_or_default(account, 'chatbot_out_of_scope_reply', DEFAULT_OUT_OF_SCOPE_REPLY)
+  end
+
+  def call_mode(account)
+    configured = account.custom_attributes['chatbot_call_mode'].to_i
+    [1, 2].include?(configured) ? configured : 1
+  end
+
+  # Apply a params hash of allowed keys to the account's custom_attributes and persist.
+  # Returns the resulting config hash. Ignores/merges only known keys.
+  def update!(account, params)
+    attrs = account.custom_attributes || {}
+    KEYS.each do |key|
+      attrs[key] = params[key] if params.key?(key)
+    end
+    account.update!(custom_attributes: attrs)
+    config(account)
+  end
+
+  # @return [Hash] full current config
+  def config(account)
+    {
+      enabled: enabled?(account),
+      system_prompt: system_prompt(account),
+      guardrail_prompt: guardrail_prompt(account),
+      out_of_scope_reply: out_of_scope_reply(account),
+      call_mode: call_mode(account)
+    }
+  end
+
+  def value_or_default(account, key, default)
+    value = account.custom_attributes[key]
+    value.presence || default
+  end
+end

app/services/chatbot/decision_service.rb

@@ -0,0 +1,206 @@
+# Chatbot decision service for the self-contained OSS chatbot.
+#
+# Given an inbound user message (plus history and retrieved knowledge), decides what the
+# bot should do:
+#   :answer  — in-scope; `answer` is the reply (grounded in the knowledge base)
+#   :refuse  — clearly off-topic (out of guardrail scope, e.g. fortune-telling); uses the
+#              account's out-of-scope reply template
+#   :handoff — related to scope but the bot can't answer (nothing in KB / undecidable);
+#              the caller should hand off to a human (conversation.bot_handoff!)
+#
+# Default: ONE LLM call returns a structured decision { decision, reason, answer }.
+# Optional TWO-call mode: a guardrail call decides in/out of scope, then an answer call
+# composes the reply from KB. Selected per account (config option; 2-call is for LLMs that
+# handle the compound single-call poorly).
+#
+# Fail-closed: no LLM credential -> { disabled: true } (never sends chat content when
+# disabled). Mirrors the Analytics::WeeklyPersonaEvaluator / Llm::AnalyticsClassifier pattern.
+module Chatbot::DecisionService
+  DECISION_SCHEMA = {
+    type: 'object',
+    additionalProperties: false,
+    properties: {
+      decision: {
+        type: 'string',
+        enum: %w[in_scope refuse handoff],
+        description: "in_scope = answer from the knowledge base; refuse = clearly off-topic and must not be answered; handoff = related to scope but bot cannot answer -> hand to a human."
+      },
+      reason: { type: 'string', description: 'One sentence justifying the decision.' },
+      answer: { type: 'string', description: 'The bot reply. Populated for in_scope; may be blank for refuse/handoff.' }
+    },
+    required: %w[decision reason answer]
+  }.freeze
+
+  Result = Struct.new(:action, :answer, :reason, :disabled, :error, keyword_init: true) do
+    def disabled? = disabled == true
+    def success? = error.nil?
+    def answer? = action == :answer
+    def refuse? = action == :refuse
+    def handoff? = action == :handoff
+  end
+
+  module_function
+
+  # @param account [Account]
+  # @param message [String] the inbound user text
+  # @param history [Array<Hash>] [{ role: 'user'|'assistant', content: String }]
+  # @param knowledge [Array<Hash>] [{ title:, content:, score: }] retrieved KB context
+  # @param call_mode [Integer] 1 (default) or 2
+  # @param system_prompt [String] optional per-account persona/system instructions
+  # @param guardrail_prompt [String] optional per-account allowed-scope instructions
+  # @return [Chatbot::DecisionService::Result]
+  def decide(account:, message:, history: [], knowledge: [], call_mode: 1, system_prompt: nil, guardrail_prompt: nil)
+    credential = Llm::Resolver.resolve(account)
+    return disabled_result if credential.nil?
+
+    # Prompts are threaded as explicit args (not module instance vars) so concurrent
+    # requests can never bleed one account's person/system prompt into another.
+    system = system_prompt.presence || SYSTEM_PROMPT
+    guardrail = guardrail_prompt.presence || GUARDRAIL_SCOPE
+
+    if call_mode == 2
+      decide_two_call(credential, message, history, knowledge, system, guardrail)
+    else
+      decide_one_call(credential, message, history, knowledge, system, guardrail)
+    end
+  rescue StandardError => e
+    Rails.logger.error("[ChatbotDecision] account=#{account&.id} #{e.class}: #{e.message}")
+    Result.new(error: e.message)
+  end
+
+  # -- 1-call mode ------------------------------------------------------------
+
+  def decide_one_call(credential, message, history, knowledge, system_prompt, guardrail_prompt)
+    response = call_llm(credential, build_one_call_prompt(message, history, knowledge, guardrail_prompt), system_prompt)
+    return Result.new(error: response[:error] || 'completion failed') if response[:error]
+
+    parsed = JSON.parse(sanitize_json(response[:content]))
+    action = normalize_action(parsed['decision'])
+    Result.new(
+      action: action,
+      reason: parsed['reason'].to_s,
+      answer: parsed['answer'].to_s,
+      disabled: false
+    )
+  rescue JSON::ParserError, TypeError
+    Result.new(error: 'LLM returned an unparsable decision')
+  end
+
+  # -- 2-call mode ------------------------------------------------------------
+
+  def decide_two_call(credential, message, history, knowledge, system_prompt, guardrail_prompt)
+    guardrail = call_llm(credential, build_guardrail_prompt(message, guardrail_prompt), system_prompt)
+    return Result.new(error: guardrail[:error] || 'guardrail failed') if guardrail[:error]
+
+    parsed = JSON.parse(sanitize_json(guardrail[:content]))
+    decision = parsed['decision']&.to_s
+    return Result.new(action: :refuse, reason: parsed['reason'].to_s, answer: '', disabled: false) if decision == 'refuse'
+    # refuse / handoff_unknown / anything-but-in_scope -> hand to a human
+    return Result.new(action: :handoff, reason: parsed['reason']&.to_s, answer: '', disabled: false) unless decision == 'in_scope'
+
+    answer_response = call_llm(credential, build_answer_prompt(message, history, knowledge), system_prompt)
+    return Result.new(error: answer_response[:error] || 'answer failed') if answer_response[:error]
+
+    Result.new(action: :answer, answer: answer_response[:content].to_s, reason: 'in_scope', disabled: false)
+  rescue JSON::ParserError, TypeError
+    Result.new(error: 'LLM returned an unparsable decision')
+  end
+
+  # -- LLM + prompt helpers ---------------------------------------------------
+
+  def call_llm(credential, prompt, system_prompt)
+    Llm::Config.with_api_key(credential[:api_key], api_base: credential[:api_base]) do |context|
+      chat = context.chat(model: MODEL).with_schema(DECISION_SCHEMA)
+      chat.with_instructions(system_prompt)
+      { content: chat.ask(prompt).content }
+    end
+  rescue StandardError => e
+    Rails.logger.error("[ChatbotDecision] LLM call failed #{e.class}: #{e.message}")
+    { error: e.message }
+  end
+
+  MODEL = Llm::Config::DEFAULT_MODEL
+
+  def build_one_call_prompt(message, history, knowledge, guardrail_prompt)
+    [
+      'Decide how the customer-service bot should respond to the customer message.',
+      '',
+      '## Guardrail scope',
+      guardrail_prompt,
+      '',
+      '## Knowledge base (retrieved, most relevant first)',
+      knowledge_text(knowledge).presence || '(no relevant knowledge found)',
+      '',
+      '## Conversation history',
+      history_text(history).presence || '(no prior messages)',
+      '',
+      "## Latest customer message\n#{message}",
+      '',
+      'If the message is in scope AND relevant knowledge exists, return decision=in_scope with the best answer grounded in the knowledge. If it is clearly outside the guardrail scope (e.g. fortune-telling, off-topic), return decision=refuse (answer may be blank). If it is related to scope but there is no knowledge to answer with, return decision=handoff.'
+    ].join("\n")
+  end
+
+  def build_guardrail_prompt(message, guardrail_prompt)
+    [
+      'You are a safety guardrail. Decide whether this customer message is within the allowed scope.',
+      '',
+      guardrail_prompt,
+      '',
+      "## Customer message\n#{message}",
+      '',
+      'Return decision: refuse if clearly outside scope; in_scope if within scope but may need knowledge to answer; handoff_unknown if related but ambiguous.'
+    ].join("\n")
+  end
+
+  def build_answer_prompt(message, history, knowledge)
+    [
+      'You are a helpful customer-service assistant for this business. Answer the customer using ONLY the provided knowledge base; do not invent facts.',
+      '',
+      '## Knowledge base',
+      knowledge_text(knowledge).presence || '(no relevant knowledge found)',
+      '',
+      '## Conversation history',
+      history_text(history).presence || '(no prior messages)',
+      '',
+      "## Customer message\n#{message}"
+    ].join("\n")
+  end
+
+  def knowledge_text(knowledge)
+    Array(knowledge).map { |k| "- #{k[:title]}: #{k[:content]}".strip }.join("\n")
+  end
+
+  def history_text(history)
+    Array(history).map { |h| "#{h[:role].to_s.capitalize}: #{h[:content]}" }.join("\n")
+  end
+
+  def normalize_action(decision)
+    case decision&.to_sym
+    when :in_scope then :answer
+    when :refuse then :refuse
+    when :handoff then :handoff
+    else :handoff
+    end
+  end
+
+  def sanitize_json(content)
+    content.to_s.gsub('```json', '').gsub('```', '').strip
+  end
+
+  def disabled_result
+    Result.new(action: nil, answer: '', reason: '', disabled: true)
+  end
+
+  SYSTEM_PROMPT = <<~PROMPT.freeze
+    You decide and then answer for a customer-service chatbot. Stay within the allowed
+    guardrail scope and be truthful and helpful. Return only the JSON object described by
+    the schema — no extra text.
+  PROMPT
+
+  GUARDRAIL_SCOPE = <<~SCOPE.freeze
+    The bot answers questions about this business's PRODUCTS, SERVICES, and related
+    support topics only. It must NOT answer unrelated or off-topic requests (e.g. personal
+    advice, fortune-telling, horoscopes, unrelated general knowledge, or any topic outside
+    the listed products/services/support).
+  SCOPE
+end

app/services/chatbot/knowledge_retriever.rb

@@ -0,0 +1,86 @@
+# KnowledgeBase retriever for the self-contained OSS chatbot.
+#
+# Given a user message, returns the top-k most relevant KnowledgeBaseFaq entries for
+# that account, ordered best-first. Retrieval is keyword-based (pg_trgm similarity on
+# title+content + topic_tag match) so it works without an embedding backend.
+#
+# A `query_embedding` param is accepted as a future extension point for embedding-fusion
+# (hybrid keyword + vector ranking), to be wired when an embedding provider is configured
+# (see plan). Returns [{ faq:, score: Float }] — the caller injects these into the LLM prompt.
+module Chatbot::KnowledgeRetriever
+  DEFAULT_LIMIT = 5
+
+  module_function
+
+  # @param account [Account]
+  # @param query [String] the user message
+  # @param query_embedding [Array<Float>, nil] reserved; embedding-fusion is a later step
+  # @param limit [Integer]
+  # @return [Array<Hash>] [{ faq:, score: Float }]
+  def retrieve(account:, query:, query_embedding: nil, limit: DEFAULT_LIMIT)
+    return [] if query.blank?
+
+    scores = score_candidates(account, query)
+    return [] if scores.empty?
+
+    max = scores.values.max
+
+    scores.map { |faq, score| { faq: faq, score: (score / max).round(4) } }
+          .sort_by { |h| -h[:score] }
+          .first(limit)
+  end
+
+  # Rank candidate FAQ entries by pg_trgm similarity + topic-tag match.
+  # @return [Hash{KnowledgeBaseFaq => Float}]
+  def score_candidates(account, query)
+    scores = {}
+    candidates(account, query).each do |faq|
+      s = faq_title_similarity(faq, query)
+      s = [s, faq_content_similarity(faq, query)].max
+      s += 0.2 if topic_match?(faq, query)
+      scores[faq] = s if s.positive?
+    end
+    scores
+  end
+
+  # Candidate set: entries whose title or content is likely relevant (pre-filter via
+  # pg_trgm word_similarity to keep the scoring pass small). Falls back to all account
+  # FAQs if pre-filter isn't available (plain AR without pg_trgm search string).
+  def candidates(account, query)
+    relation = account.knowledge_base_faqs
+    column = %(GREATEST(word_similarity(title, #{quote(query)}), word_similarity(content, #{quote(query)})))
+    relation.where("#{column} > 0.1").limit(50).to_a
+  rescue StandardError
+    relation.limit(200).to_a
+  end
+
+  def faq_title_similarity(faq, query)
+    pg_similarity(faq.title, query)
+  end
+
+  def faq_content_similarity(faq, query)
+    pg_similarity(faq.content, query)
+  end
+
+  # Token-overlap similarity ratio computed in Ruby (downcase → split → overlap / max size).
+  # Deterministic and DB-free; used to rank the small candidate pool from `candidates`.
+  def pg_similarity(text_a, text_b)
+    return 0.0 if text_a.blank? || text_b.blank?
+
+    a = text_a.downcase.split(/\s+/).reject(&:blank?)
+    b = text_b.downcase.split(/\s+/).reject(&:blank?)
+    return 0.0 if a.empty? || b.empty?
+
+    overlap = (a & b).size
+    overlap.to_f / [a.size, b.size].max.to_f
+  end
+
+  def topic_match?(faq, query)
+    q = query.downcase
+    faq.topic_tag_list.any? { |t| q.include?(t.downcase) }
+  end
+
+  def quote(value)
+    ActiveRecord::Base.sanitize_sql_like(value.to_s).gsub("'", "''")
+  end
+end

app/services/knowledge_base/import_service.rb

@@ -0,0 +1,131 @@
+# Imports markdown FAQ content into KnowledgeBaseFaq for an account (OSS self-contained
+# chatbot, phase: KB). Accepts:
+#   - pasted markdown text (content:) — split into per-heading sections
+#   - an uploaded .md / .csv / .xlsx file (file_path: + filename:)
+#
+# For markdown: each `#`/`##`/`---` section becomes one KnowledgeBaseFaq row, which
+# is the retrieval unit (title = heading, content = section body). topic_tags come
+# from explicit front-matter tags or fall back to the first heading words.
+#
+# Returns a Hash: { imported:, updated:, errors: [{ line, message }] }.
+class KnowledgeBase::ImportService
+  FRONT_MATTER_TAGS = /\A---\s*\ntags:\s*(.+?)\n---\s*\n/im
+  # Matches markdown headings (#, ##, ... up to ######). Uses [ # ]{1,6} to avoid
+  # any #{ } interpolation ambiguity in the regex literal.
+  HEADING = /^[#]{1,6}\s+(.+)$/i
+
+  def self.import(account:, content: nil, file_path: nil, filename: nil)
+    new(account: account, content: content, file_path: file_path, filename: filename).import
+  end
+
+  def initialize(account:, content: nil, file_path: nil, filename: nil)
+    @account = account
+    @content = content.to_s
+    @file_path = file_path
+    @filename = filename
+  end
+
+  def import
+    sections = @file_path ? sections_from_file : sections_from_text(@content)
+    upsert_sections(sections)
+  end
+
+  private
+
+  def sections_from_file
+    ext = File.extname(@filename.presence || @file_path.to_s).delete('.').downcase
+    text =
+      case ext
+      when 'md'
+        File.read(@file_path)
+      when 'csv'
+        require 'csv'
+        # one row per column -> naive title/content pair
+        CSV.read(@file_path).map { |r| "#{r[0]}\n\n#{r[1..].join(' ')}" }.join("\n\n")
+      when 'xlsx'
+        read_xlsx(@file_path)
+      else
+        File.read(@file_path)
+      end
+    sections_from_text(text)
+  rescue StandardError => e
+    Rails.logger.error("[KnowledgeBaseImport] parse failed: #{e.message}")
+    [{ error: "could not read file: #{e.message}" }]
+  end
+
+  def read_xlsx(path)
+    require 'roo'
+    sheet = Roo::Spreadsheet.open(path, extension: 'xlsx').sheet(0)
+    rows = (1..sheet.last_row).filter_map do |idx|
+      r = (1..sheet.last_column).map { |c| sheet.cell(idx, c).to_s }
+      "#{r[0]}\n\n#{r[1..].join(' ')}" unless r.all?(&:blank?)
+    end
+    rows.join("\n\n")
+  end
+
+  # Split markdown into per-heading sections; content preceding the first heading is
+  # treated as a single section with a derived title. A leading front-matter block
+  # (`---\ntags: ...\n---`) is stripped and its tags applied to every section.
+  def sections_from_text(text)
+    body, tags = extract_front_matter_tags(text.to_s)
+
+    sections = []
+    current = { title: nil, body: [] }
+
+    body.strip.split(/\r?\n/).each do |line|
+      if (m = line.match(HEADING))
+        # Flush the current section (even a title-less intro block) before a heading
+        sections << close_section(current, tags)
+        current = { title: m[1].strip, body: [] }
+      else
+        current[:body] << line
+      end
+    end
+    sections << close_section(current, tags)
+
+    sections.compact.reject { |s| s[:content].blank? }
+  end
+
+  # Returns [body_without_front_matter, tags_array]
+  def extract_front_matter_tags(text)
+    if (m = text.match(FRONT_MATTER_TAGS))
+      [text.sub(m[0], ''), m[1].split(/[,;\s]+/).map(&:strip).reject(&:blank?)]
+    else
+      [text, []]
+    end
+  end
+
+  def close_section(section, tags)
+    body = section[:body].join("\n").strip
+    return nil if body.blank? && section[:title].blank?
+
+    { title: section[:title].presence || body.lines.first.to_s.strip[0..80], content: body, topic_tags: tags }
+  end
+
+  def upsert_sections(sections)
+    imported = 0
+    updated = 0
+    errors = []
+    sections.each_with_index do |sec, idx|
+      next unless sec.is_a?(Hash) && sec[:content].present?
+      next if sec[:content].blank?
+
+      attrs = {
+        content: sec[:content],
+        topic_tags: sec[:topic_tags] || [],
+        source_filename: @filename
+      }
+      existing = @account.knowledge_base_faqs.find_by(title: sec[:title])
+      if existing
+        existing.update!(attrs)
+        updated += 1
+      else
+        @account.knowledge_base_faqs.create!(attrs.merge(title: sec[:title]))
+        imported += 1
+      end
+    rescue ActiveRecord::RecordInvalid => e
+      errors << { line: idx + 1, message: e.message }
+    end
+    { imported: imported, updated: updated, errors: errors }
+  end
+end

config/integration/apps.yml

@@ -87,6 +87,14 @@ slack:
   hook_type: account
   allow_multiple_hooks: false
   visible_properties: ['channel_name']
+chatbot:
+  id: chatbot
+  logo: chatbot.png
+  i18n_key: chatbot
+  action: /chatbot
+  hook_type: inbox
+  allow_multiple_hooks: false
+  visible_properties: []
 dialogflow:
   id: dialogflow
   logo: dialogflow.png

config/locales/en.yml

@@ -408,6 +408,10 @@ en:
       name: 'Dialogflow'
       short_description: 'Build chatbots to handle initial queries before transferring to agents.'
       description: 'Build chatbots with Dialogflow and easily integrate them into your inbox. These bots can handle initial queries before transferring them to a customer service agent.'
+    chatbot:
+      name: 'Chatbot'
+      short_description: 'AI chatbot that answers in-scope questions from your knowledge base and hands off to agents when needed.'
+      description: 'Enable the built-in AI chatbot: it answers questions that fall within your product/service scope using your knowledge base (md/csv/excel), refuses clearly off-topic requests, and automatically hands related-but-unanswerable conversations to a human agent.'
     google_translate:
       name: 'Google Translate'
       short_description: 'Automatically translate customer messages for agents.'

config/routes.rb

@@ -544,6 +544,27 @@ Rails.application.routes.draw do
               get :grouped_conversation_metrics
             end
           end
+          resources :analytics_reports, only: [] do
+            collection do
+              get :summary
+              get :drilldown
+              post :persona_evaluation
+              post :persona_evaluation_deliver
+              post :persona_approval_settings
+              post :persona_decision
+            end
+          end
+          resources :product_catalog_entries, only: [:index] do
+            collection do
+              post :import
+            end
+          end
+          resources :knowledge_base_faqs, only: [:index] do
+            collection do
+              post :import
+            end
+          end
+          resource :chatbot_config, only: %i[show update]
         end
       end
     end

db/migrate/20260820000000_create_knowledge_base_faqs.rb

@@ -0,0 +1,28 @@
+# KnowledgeBaseFaq — OSS-level markdown FAQ knowledge entries for the self-contained
+# chatbot. Account-scoped. Supports keyword retrieval via pg_trgm index on title and
+# an optional pgvector embedding column (populated by a later embedding step).
+#
+# This is intentionally SEPARATE from Enterprise-only Captain tables
+# (captain_assistant_responses / captain_assistants) so it works without EE.
+class CreateKnowledgeBaseFaqs < ActiveRecord::Migration[7.1]
+  def change
+    create_table :knowledge_base_faqs do |t|
+      t.bigint :account_id, null: false
+      t.string :title, null: false            # short topic/question label
+      t.text :content, null: false            # markdown body / answer
+      t.jsonb :topic_tags, default: [], null: false  # e.g. ["shipping", "refund"]
+      t.string :source_filename               # original md filename (optional)
+      t.vector :embedding, limit: 1536        # pgvector embedding (nullable; set later)
+      t.timestamps
+    end
+
+    add_index :knowledge_base_faqs, :account_id
+    add_index :knowledge_base_faqs, [:account_id, :title]
+    # pg_trgm GIN index for fuzzy keyword search on title + content
+    # (matches repo convention: gin + gin_trgm_ops, e.g. index_messages_on_content)
+    add_index :knowledge_base_faqs, :title, using: :gin, opclass: :gin_trgm_ops, name: 'index_kbf_on_title_trgm'
+    add_index :knowledge_base_faqs, :content, using: :gin, opclass: :gin_trgm_ops, name: 'index_kbf_on_content_trgm'
+    # pgvector ivfflat index for embedding similarity search (only when embeddings exist)
+    add_index :knowledge_base_faqs, :embedding, using: :ivfflat, opclass: :vector_cosine_ops, name: 'index_kbf_on_embedding'
+  end
+end

lib/integrations/chatbot/processor_service.rb

@@ -0,0 +1,107 @@
+# Chatbot processor that answers inbound chats in-place using the OSS self-contained
+# chatbot (guardrail + knowledge base + Llm::Resolver), instead of an external bot.
+#
+# Selected per inbox via an Integrations::Hook with app_id == 'chatbot' (see HookJob).
+# Mirrors Integrations::Dialogflow::ProcessorService / Integrations::Captain::ProcessorService
+# by subclassing Integrations::BotProcessorService and overriding get_response.
+#
+# Flow (from BotProcessorService#process_content):
+#   get_response(session_id, content)
+#     -> retrieve knowledge (keyword top-k)
+#     -> build history
+#     -> Chatbot::DecisionService.decide (1 or 2 calls)
+#     -> return an action marker consumed by process_response
+#   process_response(message, decision)
+#     -> :answer  -> create outbound reply (knowledge-grounded)
+#     -> :refuse  -> create outbound reply using account's out-of-scope template
+#     -> :handoff -> conversation.bot_handoff! (release bot, pending -> open, human takes over)
+#
+# Fail-closed: if the LLM is disabled or errors, we hand off to a human (safe default)
+# rather than silently not replying. Never sends chat content to the LLM when disabled.
+class Integrations::Chatbot::ProcessorService < Integrations::BotProcessorService
+  pattr_initialize [:event_name!, :hook!, :event_data!]
+
+  HANDOFF = 'chatbot_handoff'.freeze
+
+  private
+
+  # BotProcessorService calls get_response(source_id, content) then process_response.
+  # We return a Chatbot::DecisionService::Result (or a String marker for handoff) and
+  # handle rendering in process_response.
+  def get_response(_session_id, message_content)
+    return HANDOFF if message_content.blank?
+    # Fail-closed gate: if the bot is not enabled for this account, hand to a human.
+    return HANDOFF unless Chatbot::ConfigService.enabled?(conversation.account)
+
+    result = Chatbot::DecisionService.decide(
+      account: conversation.account,
+      message: message_content,
+      history: build_history,
+      knowledge: knowledge_for(message_content),
+      call_mode: chatbot_call_mode,
+      system_prompt: Chatbot::ConfigService.system_prompt(conversation.account),
+      guardrail_prompt: Chatbot::ConfigService.guardrail_prompt(conversation.account)
+    )
+
+    # Fail-closed: disabled, or any error -> human handoff (safe default).
+    return HANDOFF if result.disabled? || !result.success?
+
+    result
+  end
+
+  def process_response(message, decision)
+    return create_conversation(message, { content: out_of_scope_reply }) if decision.refuse?
+
+    if decision.handoff? || decision == HANDOFF
+      message.conversation.bot_handoff!
+      return
+    end
+
+    return if decision.answer.blank? # nothing to say
+
+    create_conversation(message, { content: decision.answer })
+  end
+
+  # -- context helpers --------------------------------------------------------
+
+  def conversation
+    @conversation ||= event_data[:message].conversation
+  end
+
+  def knowledge_for(content)
+    Chatbot::KnowledgeRetriever.retrieve(account: conversation.account, query: content)
+  end
+
+  def build_history
+    # Lightweight: the most recent few incoming/outgoing text messages (excluding this one).
+    conversation.messages
+                .where(message_type: %i[incoming outgoing], content_type: 'text')
+                .where.not(id: event_data[:message].id)
+                .order(created_at: :asc)
+                .last(8)
+                .map do |m|
+      { role: m.message_type == 'outgoing' ? 'assistant' : 'user', content: m.content.to_s }
+    end
+  end
+
+  def chatbot_call_mode
+    Chatbot::ConfigService.call_mode(conversation.account)
+  end
+
+  def out_of_scope_reply
+    Chatbot::ConfigService.out_of_scope_reply(conversation.account)
+  end
+
+  def create_conversation(message, content_params)
+    return if content_params.blank? || content_params[:content].blank?
+
+    conv = message.conversation
+    conv.messages.create!(
+      content_params.merge(
+        message_type: :outgoing,
+        account_id: conv.account_id,
+        inbox_id: conv.inbox_id
+      )
+    )
+  end
+end