IA Vendedora
Connect messages, knowledge and service state.
The review confirms routes, context retrieval and scheduling in source. No customer messages were sent and no commercial gains were measured. The account is not an end-to-end test.
Problem and proposal
IA Vendedora explores a service assistant combining messages, documents and conversation state. The challenge is integrating an external channel with persisted context and structured responses rather than calling a model without history.
The project is experimental. Sales and follow-up flows demonstrate technical integration; this page does not present unmeasured conversion, revenue or commercial-effectiveness results.
API, webhooks and responsibilities
Express organizes WhatsApp and dashboard routes, controllers, persistence and AI modules. Webhooks receive channel events; message sending and conversation querying are distinct operations. Middleware handles validation, logging and rate limiting.
The raw webhook body is preserved for signature checking. Implementation must be evaluated against the provider’s actual contract: HMAC code alone does not establish complete compatibility or audited security.
PDFs, chunks and context retrieval
The knowledge module extracts PDF text and creates overlapping chunks. Gemini generates embeddings; persistence integration organizes fragments for similarity queries, separating document preparation from contextual use during interaction.
Processing uses one message embedding to query conversation history and the document knowledge base. Results form prompt context. This retrieval path differs from StratHUB’s current direct chunk submission.
Memory and structured responses
Prompt preparation combines service stage, recent history, retrieved knowledge and long-term lead memory. Code extracts relevant information to update state and interprets model responses.
Structured output helps separate user-facing text from follow-up signals. Memory and instructions need privacy, retention and adversarial-input evaluation; stored context does not make every response correct.
Persistence, conversations and dashboard
Supabase integration maintains conversations, messages, states and analyses. Controllers query history and allow closing conversations. Dashboard routes aggregate funnel, engagement, objection and service-time data.
These routes make the flow inspectable but do not demonstrate real sales history. This page describes implemented entities and calculations without reproducing conversations, lead data or private configuration.
Scheduling and lifecycle control
FollowUpManager and FollowUpScheduler organize periodic checks and follow-up actions. node-cron controls recurrence; start and stop delimit scheduler lifecycle. Follow-up depends on persisted conversation state.
Multi-instance deployment would need execution coordination and deduplication to avoid repeated actions. Quotas, consent and channel policies must participate in the operational design.
Failures, dependencies and verification
Processing handles context-retrieval failures and may continue without vectorContext. This degradation preserves a response path but needs operator visibility so it is not mistaken for guaranteed document grounding.
The repository contains integration test and diagnostic scripts. Review inspected paths and contracts without sending messages, consuming paid models or validating assistant accuracy rates.
Skills and evolution
The project demonstrates Node.js, APIs, webhooks, embeddings, RAG components, memory, persistence and scheduling. Its value lies in connecting multiple dependencies to a stateful service flow.
Next steps include stronger route authorization, signature validation, response evaluation and scheduler coordination. This is an authorized private-project summary without source or internal data.
Private repositories. Authorized summaries; source code is not distributed here.