Develop a ready-made n8n workflow for an AI support agent for a SaaS service via the Telegram Bot API.

Hello!
We have experience in creating AI agents, RAG systems, and automation through n8n.

We can implement a ready-made AI support workflow for Venta-CRM:
— Telegram Bot API,
— OpenAI API,
— Google Drive + Google Sheets,
— vector search in the knowledge base,
— escalation flow for managers,
… — logging statistics,
— message buffering for 45 seconds.

What will be implemented:
- AI search in Google Docs through embeddings
- a separate workflow for indexing the knowledge base
- processing messages in parts
- responses only when confidence ≥ 80%
- forwarding to managers in case of no exact answer
- recording KPI and statistics in Google Sheets
- support for Telegram PM + groups
- ready JSON workflows for import into n8n
- instructions for credentials and launching

For the vector store, we recommend Qdrant – optimal for self-hosted n8n on Hetzner.

The 45-second buffer will be implemented by accumulating message context by Telegram ID with delayed processing in n8n.

Estimated timeline:
MVP — 3–5 days
Full production version — 7–10 days.

We would be happy to discuss the details and implement a ready-made working solution for you 🙂

Show original

Translate

I see that you have not just an "AI bot," but a normal first-line support task: Telegram → 45-second message buffer → search in the indexed knowledge base → respond only on exact match → escalate to managers if confidence is below threshold.

We are currently implementing similar logic for AI managers in messengers. I wouldn't read Google Docs for every request — it would be more appropriate to create a separate indexing: Google Docs → chunks → embeddings → vector store, and then search in the vector base at runtime.

I would implement the 45 seconds as a debounce at the backend level: all user messages are temporarily collected by chat_id/user_id, each new message resets the timer, and only after 45 seconds of silence does the agent process the request as one complete context.

If n8n JSON is a strict requirement — this is important to clarify. But I can offer the same logic as a separate stable backend service with Telegram Bot API, OpenAI, vector search, escalation to a group of managers, and logging statistics. Architecturally, this will be more scalable and controllable than a complex workflow in n8n.

Message me privately, and I will consult you in more detail and show you how it already works for us.

Show original

Translate

Hello! I can complete this bot in 10 days, the price will be $200. If you are ready, we can collaborate.

Show original

Translate

Pavlo, we have communicated before, not very successfully, but I still hope my Workflow was useful to you.

I am interested in this project, I see nothing impossible or difficult for myself, I have experience with all the listed tools.

1. Have you created similar AI support agents?
Not in this exact format, but I don't see anything complicated.

2. Which vector store do you recommend for this task?
Supabase + pgvector
…
3. How do you plan to implement the 45-second buffer?
There are many options, here is the most logical one for me right now:
When a person writes, we log their request in the database, we also log 2 variables:
- last_message_time: date and time of the last message
- agent_status: active/inactive
Then the condition is, if agent_status = inactive, a script is triggered.
Next, we set a wait node for 45 seconds, after which we check last_message_time that was initially recorded against what is currently in the database; if it differs, it means we have received a new message, so we go back and wait another 45 seconds.

Only when the message time does not change during this period do we proceed with the script.
I hope I explained it clearly)

4. How will you determine confidence >= 80%?
- Option 1: Add an instruction in the prompt on how to determine it and ask the agent to provide a response already with this variable's indicator.
- Option 2: Add another agent that will check the response of the previous agent against what is in the database.

5. Estimated completion time:
5-7 days.

Show original

Translate

Nice brief, interesting task.
I am currently working in this direction, my stack is perfectly suited for the task.
I actively use Gemini Pro 3.1, it is tailored for similar tasks.
I have my own server - I will test everything.

I suggest using the GPT-4o-mini model, it is very cheap and smart enough to handle this task.

If you do not choose me as the executor, I will still implement this project, because, I repeat, the task is quite interesting, and the brief is of the highest level.

… Have a nice day.

Show original

Translate

Good day, I am writing on behalf of the company Devoxen. We specialize in AI automation and workflow solutions based on n8n, Telegram Bot API, and OpenAI. We have extensive experience in building AI support agents with RAG architecture, integrating Google Drive/Sheets, and escalation logic to managers without "inventing" responses. For your case, we would recommend a vector store like Qdrant or Pinecone — for n8n, this is a stable and scalable option for future knowledge base expansion. We will implement a 45-second buffer through Redis/Data Store with an aggregation message window, so the bot correctly combines messages into one request. Confidence will be determined in a combined manner: similarity score embeddings + additional AI verification of response relevance before sending to the user. We will also lay the architecture for future scaling to multiple bots and separate clients.

We can proceed without unnecessary questions and time expenditures. We also provide a guarantee and support if desired. We will be able to take on your project immediately after discussing the technical specifications and the structure of the knowledge base.

I suggest moving to personal messages for a more detailed conversation.

Show original

Translate

Hello!

1. I have experience with similar AI support agents - I have worked with Telegram Bot API, OpenAI API, Google Drive/Sheets, and RAG logic.

2. For the vector store, I recommend Pinecone or Qdrant - both integrate well with n8n and OpenAI embeddings.

3. A 45-second buffer - I will implement this through the n8n Wait node with message accumulation before processing.

4. Confidence >= 80% - I determine this through cosine similarity when searching in the vector store + additional verification through GPT to see if the response matches the query.
…
5. Timeline - 7-10 days.

I am ready to discuss the details and pricing.

Show original

Translate

Вітаю! Завдання зрозуміле, це вже повноцінний AI support agent з RAG-архітектурою для n8n, а не просто Telegram-бот. Можу реалізувати готовий production-ready workflow під ваш сценарій.

Маю практичний досвід з:
— n8n self-hosted
— Telegram Bot API
— OpenAI API
— RAG / embeddings / vector search
— Google Drive & Google Sheets integrations
— AI support workflows
… — escalation systems
— multi-step conversational logic

Як бачу реалізацію:

Telegram → n8n → message buffer → AI classification → vector search → confidence validation → response/escalation.

Що рекомендую:
— OpenAI embeddings text-embedding-3-small
— vector store: Qdrant або Supabase pgvector
— GPT-4.1 mini / GPT-4o mini для відповіді
— окремий indexing workflow
— окремий support workflow
— Google Sheets для KPI/logging

Як реалізую buffer 45 секунд:
— Redis або n8n data store/session aggregation
— накопичення повідомлень по Telegram user ID
— debounce logic перед запуском AI flow

Як визначатиму confidence:
— similarity score vector search + AI verification layer
— threshold >= 0.80
— додаткова перевірка “чи є відповідь повністю підтвердженою knowledge base”

Важливий момент:
агент не буде hallucinate/вигадувати відповіді — тільки grounded answers із knowledge base.

Що буде в результаті:
— готовий JSON workflow support agent
— окремий JSON workflow indexing
— інструкція по credentials
— інструкція по Telegram/OpenAI/Google Drive
— Google Sheets statistics setup
— mapping variables
— tested flow у staging середовищі

Також можу закласти архітектуру під:
— multi-client support
— multi-tenant bots
— future WhatsApp/website integration
— manager dashboard/escalation layer
— feedback loop для knowledge base growth

Можу також запропонувати production architecture для Hetzner + Docker deployment + backup/recovery flow для n8n.

Hello! Very mature technical specification, especially the confidence gate of 80% and the 45-second buffer. This is a production support agent, not just a bot with RAG.

I will respond point by point:

1. I regularly build AI support agents with RAG — Telegram + embeddings + vector store + escalation.
2. Vector store: Supabase pgvector. For 50+ documents, self-hosted and free, natively with Hetzner.
3. 45-second buffer: Wait node + session_id in Supabase as a queue, on timeout concat all messages with the same chat_id.
4. Confidence ≥80%: combined signal — cosine similarity + a separate decision step where the model assesses the relevance of the chunk (1–10). This protects against hallucinations more reliably than a single similarity score.
5. Deadline: 12 days with full testing on a real bot before handing over JSON.
…
Background: MSc in Strategic PM (Lazarski), PRINCE2, 4 years of PM — helps not to falter on acceptance criteria and documentation. My architecture is very close to the pipeline (multi-source + AI scoring + structured output) and many others in my profile.

Show original

Translate

Good day.
I have reviewed the specifications. Special thanks for the detailed specifications.

1) I have an example of searching for an AI agent in the company's knowledge base (information about the company, frequently asked questions, details about job vacancies).
A person can ask the bot about the company or details regarding any of the vacancies, and the bot responds based on the information in the knowledge base in the Google document. I can send a video overview privately.

2) I suggest using Pinecone as a vector database for chunking and RAG search (in addition to chunks, labels like /company /vacancies, etc. can be assigned for faster searching).

3) When an incoming message is received in the workflow, add saving to the database and a 45-second delay. If a new message comes in, the agent will check the timestamp; if the time between messages is within 45 seconds, it will summarize the last messages, and there will be one message for the agent to process.
…
4) This still needs to be thought through.
5) 12 days (including testing).
6) I can send an example privately.

Show original

Translate

I have created similar AI support agents on n8n — Telegram + RAG + escalation to managers, so the task is clear from start to finish.
Regarding your questions:

Yes, I have made similar agents for support — gathering requests, searching the knowledge base, passing to a manager if confidence is low.
For the vector store — I recommend Pinecone or Supabase pgvector, both integrate well with n8n.
The 45-second buffer — I will implement through the Wait node + Session ID, collecting messages into one queue.
Confidence >= 80% — through OpenAI embeddings cosine similarity score when searching in the vector store.
Timeline — 5-7 days with testing on a real bot.
I can show examples of similar workflows in private.
…
I will provide the ready JSON workflow, instructions for connecting credentials, and a list of places where you need to insert your tokens.

Show original

Translate

Welcome! Our team with 4 years of experience in engineering development and process automation is ready to create a stable, optimized, and scalable n8n workflow for your AI agent. We have deep technical expertise in bot development, data parsing, and integrating complex APIs. With knowledge of Python and JavaScript, we not only assemble standard blocks in n8n but also know how to write custom scripts (inside Code nodes) for complex data array processing, flexible AI context management, and non-standard integrations. We are ready to discuss the logic of your future agent's operation and data sources in private messages to propose the best implementation option.

Show original

Translate

Hello!
1. Have you made similar AI support agents? Directly a RAG agent in n8n — no, but I have practical experience with n8n + OpenAI API, Telegram Bot API, and automation through Google Drive/Sheets. I understand the RAG architecture and how to implement it in n8n using the built-in vector store + embeddings.

2. Which vector store do I recommend? For your scale (20–50 documents, self-hosted Hetzner) — Supabase with pgvector. It's free, integrates well with n8n, and doesn't require a separate service. Pinecone — if you want a cloud solution without administration.

3. How will I implement a 45-second buffer? Through Supabase or Google Sheets as a temporary queue: with each message from the user, I record it there with a timestamp, wait 45 seconds through the Wait node, then gather all messages with the same chat_id from the last 50 seconds and send them to the agent as one request.

4. How will I determine Confidence >= 80%? Through cosine similarity score during vector search — n8n returns this value. Additionally, I prompt the model to assess how well the found context matches the question (1–10), and I combine both signals.

… 5. Deadline: 7–10 days for a complete MVP with testing. Manager buttons (Took to work / Responded) — a separate block, I can architect it, but implement it in the second iteration to avoid delaying the delivery.

What I will provide: both JSON workflows, a step-by-step guide for connecting credentials, a description of variables, and a video demonstration of the bot's operation.

If you have any questions — feel free to write, I would be happy to discuss the details.

Show original

Translate

Hello! I will implement your workflow for Venta-CRM on n8n with RAG logic. I have experience in creating AI supports, where critical accuracy of responses without hallucinations is essential.

For the vector database, I recommend Pinecone — it is reliable and has a native node in n8n. I will implement a 45-second buffer using the Wait node with a condition to check for new messages, in order to merge them into a cohesive request. I will set the confidence at 80% through a system prompt, where the AI evaluates the relevance of the found instruction before sending.

The timeline is 14-21 days.

I am ready to discuss the details and start working.

Show original

Translate

Good day.

The specifications are detailed and clear, showing a mature approach. The task is understood: n8n workflow for the AI support agent of the first line for Venta-CRM from Telegram, RAG over Google Drive, with handover to managers and statistics.

Regarding your questions:

1. AI support agents with RAG have been done in production; this is our core type of projects. A relevant case is Winbix.AI with $2K MRR, which uses a similar architecture.

2. Vector store for this task: Qdrant Cloud (free tier up to 1GB will suffice for 50+ documents) or Pinecone starter. Qdrant is slightly better in search quality and cheaper at scale. In n8n, we connect via HTTP Request nodes to the Qdrant API.
…
3. A buffer of 45 seconds through the n8n Wait node plus Redis or n8n static data for accumulating messages. With each new message from the user, reset the timer. After 45 seconds of silence, we combine the messages into one request and trigger the AI.

4. Confidence 80%: hybrid approach. First, cosine similarity score from the vector store (must be above 0.78). Then LLM-as-judge via GPT-4o with structured output: "Is the answer fully supported by the context? Return confidence 0-100." Only if both conditions are above the threshold do we respond. Otherwise, escalation.

5. Timeline: 5-7 working days from the start. Includes indexing the workflow, the main workflow, testing on a real bot, and documentation.

6. Model: GPT-4o-mini for classification and judge, GPT-4o for generating the response. Embeddings text-embedding-3-small (cheap and quality for the Ukrainian language).

Before the main contract, we are ready to show a working prototype of the key block (45 sec buffer plus RAG search) on a test bot.

Portfolio: quentar.space/en/startups

I await your personal message.

Show original

Translate

Good day. I represent the Nexus Core team. You have a very strong and specific technical specification: you need not an "AI chatbot," but a ready production workflow for n8n that operates as the first line of support for Venta-CRM via Telegram, with a pre-indexed knowledge base on Google Drive, clear escalation logic, statistics collection, and scalability for other clients. We are just taking on such tasks where the theoretical RAG is not important, but a real working scheme: the bot does not invent, does not respond "approximately," can gather messages in parts, correctly distinguishes a simple greeting from a full request, and only forwards to managers those inquiries where there is truly no confirmed response. For this, we recommend an architecture with a separate workflow for indexing, OpenAI embeddings, a vector store divided by knowledge base, runtime search only on indexed data, and a clear confidence gate to prevent the agent from hallucinating.

From our side, we can provide exactly what you described: a ready JSON workflow for runtime, a separate JSON for indexing, instructions for credentials, Telegram Bot API, Google Drive, Google Sheets, a description of the logic, a list of variables, and a tested scenario in a test environment before handover. For the vector store for such a task, we optimally see Supabase pgvector or Qdrant: both are well-suited for self-hosted logic, but for simpler deployment and future scalability, Supabase pgvector is often more advantageous. The 45-second buffer will be implemented not "head-on," but in a way that does not break the dialogue logic and does not create chaos with several messages in a row. Confidence >= 80% will be determined not by a single raw number from the LLM, but by a combination of search relevance, chunk matching verification, and a separate decision step, so that the response goes to the client only when it is truly confirmed by the knowledge base. Cases of similar AI support flows are under NDA, but we can explain in private messages how we implemented RAG, Telegram bots, n8n automation, and escalation to managerial channels in detail.

To immediately fix the term and cost without error, it is important to clarify three things: do you already have a ready Telegram group of managers and a bot, or is that also included in the launch; do you plan for the MVP to only use Telegram, without additional channels; and do you need the buttons "Took into work / Responded / Add to knowledge base" in the first version, or should we lay them out in the architecture for the second stage? Write in private messages — we will clarify the details and provide a fixed estimate on terms, stages, and final cost.

Show original

Translate

Hello.
I am ready to take on this project.
The specifications are very detailed and clear, so I can start the development right away.
I will deploy everything on my Hetzner server (I already have one).

Show original

Translate

Hello.

I develop bots for Telegram. I'm ready to take on the task. Write to me, and we will discuss.

Show original

Translate

Hello. I have experience with n8n/Telegram bots. Ready to collaborate. Contact me!

Show original

Translate

Hello! After reviewing your project, I am ready to start working on it. Let's discuss the details for the best result.

Show original

Translate

Hello. A 45-second delay before responding is a smart move to gather a complete request, but in a group chat, it will create chaos unless logic is added to tie it to a specific user and thread. I created a similar AI agent for a logistics SaaS: embeddings from 40+ Google Docs, Pinecone as the vector store, manual + weekly re-indexing through a separate n8n workflow. The bot responded in private messages and tracked threads in the group where it was mentioned. The average response time was 3-4 seconds after the delay, with a search accuracy of about 85% on Ukrainian technical documentation. For Venta-CRM, I recommend deepseek/kimi/gemini for classification + for the final response - an optimal balance of cost and quality. Regarding scaling: it's worth incorporating the tenant_id variable at the design stage to avoid reworking the logic when connecting a new client. Which vector store are you considering - Pinecone, Weaviate, or are you ready for self-hosted Qdrant?

Show original

Translate