Add RAG-backed AI agents#428
Draft
davibritovilela52-design wants to merge 64 commits into
Draft
Conversation
Drops the BYOK AI reply-assistant feature entirely (auto-reply, knowledge base, playground, usage dashboard, /agents section, /api/ai/* routes) ahead of a from-scratch WhatsApp AI agent design. Bundled with the pending Decizyon brand identity changes that were staged alongside it. Migration required: apply supabase/migrations/037_drop_ai.sql (destructive — drops AI config, knowledge base, usage, and conversation auto-reply data).
Supersedes the two discarded plans (pipeline-stage AI classifier, NL automation builder) that were written against the now-removed BYOK AI infra. Single structured-decision-per-event architecture instead of a full agent loop, per chatwoot/OpenHands pattern review.
13 tasks: schema, provider-agnostic generateJson foundation, AiConfig settings UI, move_deal_stage/deal_stage_changed engine + builder support, moveDealStage helper, WhatsApp agent decision+dispatch, automation copilot generation+UI, full regression pass.
Implements Task 2: AI provider foundation with provider-agnostic structured-output call. Creates types, provider adapters for OpenAI and Anthropic, and the main generateJson<T>() function that handles JSON parsing with fallback for prose-wrapped responses. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a new Settings > AI agent panel that reads/writes GET and PUT /api/ai/config (Task 3), letting an account configure its AI provider/model/API key and auto-reply/pipeline-move behavior.
…al_stage condition
…er, deal_stage condition
…ng empty resources
Loads a conversation's recent text history and linked deal stage (buildAgentContext), then asks the LLM for a structured decision (reply/tags/stage move/handoff) via generateJson and sanitizes the result against the account's real tags/pipelines/stages so a hallucinated id can never reach the executor.
…ilently degrading Both messages and deals queries now destructure and check the error field, throwing descriptive errors naming the table and underlying error message when non-null, following the established pattern in src/lib/automations/resources.ts. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Wires the Task 9 decision engine into the real inbound webhook path: after the automation-trigger dispatch, fire dispatchInboundToAgent fire-and-forget so the AI decision (reply/tag/move-stage/handoff) runs without ever blocking or breaking the webhook's response to Meta. Adds the aiAgentDecision rate-limit bucket (30/min per account, checked before the AI call). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… of a racy client-computed update
Adds generateAutomationFromPrompt, the second and last consumer of the "never trust a model-returned id" sanitizer pattern in this plan: a single generateJson call per chat turn that returns either a clarifying question or a draft automation built only from real tag/pipeline/stage ids. Fixes a parent_index indexing bug present in the task brief's reference implementation: the model's parent_index values are positions in its own raw steps array, but ALLOWED_STEPS filtering can drop steps, shortening the output array and desynchronizing raw vs. output indices. Resolving parent_index against the raw loop index (as the brief's code did) could produce self- or forward-referencing parent links once an earlier raw step was dropped. Fixed with a two-pass approach: filter first, then remap parent_index through a raw-index -> output-index map, keeping only backward references. Covered by two added regression tests beyond the brief's five.
Adds POST /api/automations/generate (agent+, rate-limited) wrapping
generateAutomationFromPrompt with pre-flight activation validation,
plus an AiCopilotPanel chat dialog wired into the automations list
page ("Ask AI" button next to "New automation"). Approved drafts are
created inactive via the existing POST /api/automations and open in
the normal builder for review.
…i-turn refinement
…cap reply_text length Addresses findings from the final whole-branch review: - loadAutomationResources() no longer fetches all tenants' pipeline_stages unfiltered when called via the service-role client (agent-dispatch's hot path on every inbound WhatsApp message) — now scoped via the account's own pipeline ids. - agent-dispatch.ts's handoff conversation update now logs on failure instead of silently swallowing it, matching every other write in the same function. - agent-decide.ts's sanitize() now caps reply_text at 4096 chars (WhatsApp's own text message limit), matching the existing handoff_reason cap pattern.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Test Plan
Notes
npm run buildcompiles and passes TypeScript, then fails prerendering unchanged/automations/newwith a Next invariant (Expected workStore to be initialized). Final branch review judged that failure outside this branch's diff.038_ai_agent.sqland041_ai_knowledge_rag.sql, plus a validENCRYPTION_KEYin target environments.