mirror of
https://github.com/khoaliber/n8nworkflows.xyz.git
synced 2026-04-28 00:29:22 +00:00
creation
This commit is contained in:
nusquama
+308
@@ -0,0 +1,308 @@
|
||||
Turn Telegram chatbots into GPT-4o-mini advisors with Postgres memory and Shopify data
|
||||
|
||||
https://n8nworkflows.xyz/workflows/turn-telegram-chatbots-into-gpt-4o-mini-advisors-with-postgres-memory-and-shopify-data-12885
|
||||
|
||||
|
||||
# Turn Telegram chatbots into GPT-4o-mini advisors with Postgres memory and Shopify data
|
||||
|
||||
## 1. Workflow Overview
|
||||
|
||||
**Purpose:**
|
||||
This workflow turns a Telegram chatbot into a **GPT-4o-mini powered advisor** (“SOFÍA”, a coffee barista persona) that answers using:
|
||||
- **Persistent conversation memory** stored in **Postgres** (per Telegram chat/user),
|
||||
- **Real-time Shopify product catalog** (names + prices) to avoid hallucinations,
|
||||
- A strict **JSON response format**, validated and repaired with a fallback.
|
||||
|
||||
**Target use cases:**
|
||||
- Customer support / sales assistant that remembers user preferences over time
|
||||
- Product recommendation based on live catalog data
|
||||
- Multi-turn conversations where identity (user name) is required before advising
|
||||
|
||||
### 1.1 Input Reception & User Identification
|
||||
Receives Telegram messages and extracts a stable `chat_id` + message text for downstream nodes and memory session keying.
|
||||
|
||||
### 1.2 Real-Time Business Context (Shopify)
|
||||
Loads the current Shopify product list so the AI can quote exact product titles and prices.
|
||||
|
||||
### 1.3 AI Reasoning with Memory (OpenAI + LangChain Agent + Postgres Memory)
|
||||
Runs an AI Agent using GPT-4o-mini, injecting:
|
||||
- user message,
|
||||
- Postgres chat history,
|
||||
- Shopify product context inside the system prompt.
|
||||
|
||||
### 1.4 Output Validation & Delivery (JSON parsing + Telegram reply)
|
||||
Parses/repairs the agent output into valid JSON and sends `response` back to Telegram, with a safe fallback if parsing fails.
|
||||
|
||||
---
|
||||
|
||||
## 2. Block-by-Block Analysis
|
||||
|
||||
### Block 1 — Input Reception & User Identification
|
||||
**Overview:** Captures incoming Telegram messages and normalizes the payload to a minimal structure (`text`, `chat_id`) used throughout the workflow (including memory session keying).
|
||||
|
||||
**Nodes involved:**
|
||||
- `Telegram Trigger`
|
||||
- `Code in JavaScript`
|
||||
|
||||
#### Node: Telegram Trigger
|
||||
- **Type / role:** `n8n-nodes-base.telegramTrigger` — entry point; listens for Telegram updates.
|
||||
- **Configuration (interpreted):**
|
||||
- Updates: `message` (only standard chat messages)
|
||||
- Uses Telegram credentials: **“Telegram account”**
|
||||
- **Outputs:** Telegram update JSON (includes `message.text`, `message.chat.id`, etc.)
|
||||
- **Connections:**
|
||||
- Output → `Code in JavaScript`
|
||||
- **Potential failures / edge cases:**
|
||||
- Telegram credential invalid/revoked
|
||||
- Bot not added to chat / privacy settings
|
||||
- Non-text messages (photos, stickers) may not contain `message.text` → downstream expressions must handle empty string
|
||||
|
||||
#### Node: Code in JavaScript
|
||||
- **Type / role:** `n8n-nodes-base.code` — normalizes incoming Telegram payload.
|
||||
- **Key logic:**
|
||||
- Extracts:
|
||||
- `chatMessage = $json.message.text || ""`
|
||||
- `chatId = $json.message.chat.id || 0`
|
||||
- Returns a clean object: `{ text, chat_id }`
|
||||
- **Why this matters:**
|
||||
- Creates a guaranteed `chat_id` for Postgres memory session keys.
|
||||
- **Connections:**
|
||||
- Input: `Telegram Trigger`
|
||||
- Output → `Get many products`
|
||||
- **Potential failures / edge cases:**
|
||||
- If Telegram payload differs (e.g., edited messages), `message` can be missing → would throw unless guarded. Current code assumes `$json.message` exists.
|
||||
- If `chat_id` becomes `0`, memory will collide for all such failures (bad). Consider defensive checks.
|
||||
|
||||
---
|
||||
|
||||
### Block 2 — Real-Time Context (Shopify)
|
||||
**Overview:** Fetches the live product catalog from Shopify so the AI can list exact product names and prices.
|
||||
|
||||
**Nodes involved:**
|
||||
- `Get many products`
|
||||
|
||||
#### Node: Get many products
|
||||
- **Type / role:** `n8n-nodes-base.shopify` — reads product data from Shopify.
|
||||
- **Configuration (interpreted):**
|
||||
- Resource: `product`
|
||||
- Operation: `getAll`
|
||||
- Auth: OAuth2
|
||||
- Credentials: **“Shopify Access Token account”**
|
||||
- **Output:** An item per product (with fields like `title`, `variants[0].price`, etc.)
|
||||
- **Connections:**
|
||||
- Input: `Code in JavaScript`
|
||||
- Output → `AI Agent`
|
||||
- **Potential failures / edge cases:**
|
||||
- OAuth token expired/scopes missing (e.g., products read scope)
|
||||
- Large catalogs may paginate; “getAll” can be slow/time out depending on store size
|
||||
- Some products may have empty `variants` or missing `variants[0].price` → the system prompt expression can break or produce `undefined`
|
||||
|
||||
---
|
||||
|
||||
### Block 3 — AI Reasoning & Memory
|
||||
**Overview:** Runs the AI Agent with a strict system protocol (name capture, catalog precision rules, memory rules). It uses Postgres chat history to remember the user and preferences.
|
||||
|
||||
**Nodes involved:**
|
||||
- `OpenAI Chat Model`
|
||||
- `Postgres Chat Memory`
|
||||
- `AI Agent`
|
||||
|
||||
#### Node: OpenAI Chat Model
|
||||
- **Type / role:** `@n8n/n8n-nodes-langchain.lmChatOpenAi` — provides the LLM for the agent.
|
||||
- **Configuration (interpreted):**
|
||||
- Model: `gpt-4o-mini`
|
||||
- Credentials: **“OpenAi account”**
|
||||
- No additional tools enabled (built-in tools empty)
|
||||
- **Connections:**
|
||||
- Output (AI language model channel) → `AI Agent`
|
||||
- **Potential failures / edge cases:**
|
||||
- API key invalid / billing / rate limits
|
||||
- Model unavailable in region/account
|
||||
- Timeouts on long contexts (especially if many products + memory)
|
||||
|
||||
#### Node: Postgres Chat Memory
|
||||
- **Type / role:** `@n8n/n8n-nodes-langchain.memoryPostgresChat` — persistent chat memory backend.
|
||||
- **Configuration (interpreted):**
|
||||
- Session ID type: `customKey`
|
||||
- Session key: `{{ $('Code in JavaScript').item.json.chat_id }}`
|
||||
- Context window length: `30` (last 30 messages/interactions retained in context)
|
||||
- Credentials: **“Postgres account 2”**
|
||||
- **Connections:**
|
||||
- Output (AI memory channel) → `AI Agent`
|
||||
- **Potential failures / edge cases:**
|
||||
- Postgres connectivity/auth/SSL issues
|
||||
- Missing required tables/schema (node may auto-create depending on version/config)
|
||||
- If `chat_id` is wrong/empty, users may share memory or lose history
|
||||
|
||||
#### Node: AI Agent
|
||||
- **Type / role:** `@n8n/n8n-nodes-langchain.agent` — orchestrates prompt + memory + model to produce the response.
|
||||
- **Configuration (interpreted):**
|
||||
- User text input: `{{ $('Telegram Trigger').item.json.message.text }}`
|
||||
- Note: This bypasses the normalized `text` from `Code in JavaScript` (works for text messages, but less robust for non-text updates).
|
||||
- **System message** (Spanish) defines:
|
||||
- Persona: “SOFÍA” (Zyntha Coffee barista)
|
||||
- Mandatory greeting protocol depending on whether the user’s name is known from memory
|
||||
- “Catalog precision” rule: must list exact product names and prices from Shopify section
|
||||
- “Focus rule” for discussing a specific coffee
|
||||
- Strict JSON output schema:
|
||||
```json
|
||||
{
|
||||
"intent": "setup | recommendation | question",
|
||||
"response": "...",
|
||||
"product": "Name or null",
|
||||
"qty": null
|
||||
}
|
||||
```
|
||||
- Injected product list (dynamic expression):
|
||||
- `{{ $('Get many products').all().map(p => p.json.title + " (Precio: $" + p.json.variants[0].price + ")").join(', ') }}`
|
||||
- Output parser enabled (`hasOutputParser: true`) — expects structured output.
|
||||
- **Connections:**
|
||||
- Inputs:
|
||||
- Model from `OpenAI Chat Model`
|
||||
- Memory from `Postgres Chat Memory`
|
||||
- Main input from `Get many products` (this is how the agent “sees” the Shopify items in the execution context)
|
||||
- Output → `Code in JavaScript 2`
|
||||
- **Potential failures / edge cases:**
|
||||
- If Shopify products output is large, system prompt becomes very long → higher latency / context limits
|
||||
- If `variants[0]` missing, the system message expression may error or render invalid text
|
||||
- If the agent outputs non-JSON (or JSON with extra text), downstream parsing may fail (handled by fallback)
|
||||
|
||||
---
|
||||
|
||||
### Block 4 — Output Validation & Delivery
|
||||
**Overview:** Ensures the agent response is valid JSON and always returns a safe message if parsing fails, then replies to the Telegram user.
|
||||
|
||||
**Nodes involved:**
|
||||
- `Code in JavaScript 2`
|
||||
- `Send a text message`
|
||||
|
||||
#### Node: Code in JavaScript 2
|
||||
- **Type / role:** `n8n-nodes-base.code` — attempts to extract and parse JSON from the agent output; otherwise returns a fallback JSON object.
|
||||
- **Key logic (interpreted):**
|
||||
- Reads `items[0].json.output` (agent raw output).
|
||||
- Extracts substring between first `{` and last `}` to tolerate extra text.
|
||||
- `JSON.parse()` that substring.
|
||||
- **Fallback:** if parsing fails, returns a forced `"setup"` response asking for the user’s name.
|
||||
- **Important implementation detail in n8n:**
|
||||
- The code returns a plain object (`return parsed;` / `return { ... }`) instead of `return [{ json: parsed }]`. In many n8n versions, the Code node should return an array of items. If your n8n build is strict, you must wrap outputs as items.
|
||||
- **Connections:**
|
||||
- Input: `AI Agent`
|
||||
- Output → `Send a text message`
|
||||
- **Potential failures / edge cases:**
|
||||
- `items[0].json.output` missing (agent node output shape changes) → runtime error
|
||||
- If AI returns JSON but with trailing commas/unquoted keys → parse fails → fallback triggers
|
||||
- If fallback triggers too often, users will be stuck in “tell me your name” loop
|
||||
|
||||
#### Node: Send a text message
|
||||
- **Type / role:** `n8n-nodes-base.telegram` — sends the final response back to Telegram.
|
||||
- **Configuration (interpreted):**
|
||||
- Chat ID: `{{ $('Code in JavaScript').first().json.chat_id }}`
|
||||
- Text: `{{ $json.response }}`
|
||||
- Credentials: **“Telegram account”**
|
||||
- **Connections:**
|
||||
- Input: `Code in JavaScript 2`
|
||||
- Output: none (end)
|
||||
- **Potential failures / edge cases:**
|
||||
- If `chat_id` extraction failed earlier, message may go to chat `0` (invalid) and fail
|
||||
- Telegram message length limits (very long AI responses can fail)
|
||||
|
||||
---
|
||||
|
||||
## 3. Summary Table
|
||||
|
||||
| Node Name | Node Type | Functional Role | Input Node(s) | Output Node(s) | Sticky Note |
|
||||
|---|---|---|---|---|---|
|
||||
| Telegram Trigger | n8n-nodes-base.telegramTrigger | Entry point: receives Telegram messages | — | Code in JavaScript | ## Section 1 — Input & User Identification; Receives the incoming message and extracts the chat_id used to uniquely identify the user. This ensures every message can be linked to the correct conversation history. |
|
||||
| Code in JavaScript | n8n-nodes-base.code | Normalize payload; extract `text` and `chat_id` | Telegram Trigger | Get many products | ## Section 1 — Input & User Identification; Receives the incoming message and extracts the chat_id used to uniquely identify the user. This ensures every message can be linked to the correct conversation history. |
|
||||
| Get many products | n8n-nodes-base.shopify | Fetch live Shopify product catalog | Code in JavaScript | AI Agent | ## Section 2 — Real-Time Context; Fetches the live product catalog directly from the store. This prevents the AI from hallucinating product names or prices. |
|
||||
| OpenAI Chat Model | @n8n/n8n-nodes-langchain.lmChatOpenAi | LLM provider (gpt-4o-mini) | — (model input implicit) | AI Agent | ## Section 3 — AI Reasoning & Memory; Combines user input, historical memory, and real-time data to generate a contextual response. This is where personalization, rules, and long-term memory are applied. |
|
||||
| Postgres Chat Memory | @n8n/n8n-nodes-langchain.memoryPostgresChat | Persistent chat memory keyed by `chat_id` | — (memory input implicit) | AI Agent | ## Section 3 — AI Reasoning & Memory; Combines user input, historical memory, and real-time data to generate a contextual response. This is where personalization, rules, and long-term memory are applied. |
|
||||
| AI Agent | @n8n/n8n-nodes-langchain.agent | Orchestrates prompt + memory + model; generates JSON output | Get many products (+ model + memory channels) | Code in JavaScript 2 | ## Section 3 — AI Reasoning & Memory; Combines user input, historical memory, and real-time data to generate a contextual response. This is where personalization, rules, and long-term memory are applied. |
|
||||
| Code in JavaScript 2 | n8n-nodes-base.code | Parse/repair JSON; fallback response | AI Agent | Send a text message | ## Section 4 — Output & Delivery; Validates that the AI response is valid JSON and applies a fallback if needed. Sends the final, safe response back to the user via Telegram. / ⚠️ Important; This workflow requires valid JSON output from the AI. If you modify the system prompt, make sure the response format is preserved to avoid parsing errors. |
|
||||
| Send a text message | n8n-nodes-base.telegram | Sends final response to Telegram | Code in JavaScript 2 | — | ## Section 4 — Output & Delivery; Validates that the AI response is valid JSON and applies a fallback if needed. Sends the final, safe response back to the user via Telegram. / ⚠️ Important; This workflow requires valid JSON output from the AI. If you modify the system prompt, make sure the response format is preserved to avoid parsing errors. |
|
||||
| Sticky Note7 | n8n-nodes-base.stickyNote | Documentation/comment | — | — | ## AI Advisor with Memory and Real-Time Context; How it works… How to set up… Customization… |
|
||||
| Sticky Note8 | n8n-nodes-base.stickyNote | Documentation/comment | — | — | ## Section 1 — Input & User Identification; Receives the incoming message… |
|
||||
| Sticky Note9 | n8n-nodes-base.stickyNote | Documentation/comment | — | — | ## Section 2 — Real-Time Context; Fetches the live product catalog… |
|
||||
| Sticky Note10 | n8n-nodes-base.stickyNote | Documentation/comment | — | — | ## Section 3 — AI Reasoning & Memory; Combines user input… |
|
||||
| Sticky Note11 | n8n-nodes-base.stickyNote | Documentation/comment | — | — | ## Section 4 — Output & Delivery; Validates that the AI response… |
|
||||
| Sticky Note | n8n-nodes-base.stickyNote | Documentation/comment | — | — | ⚠️ Important; This workflow requires valid JSON output from the AI… |
|
||||
|
||||
---
|
||||
|
||||
## 4. Reproducing the Workflow from Scratch
|
||||
|
||||
1. **Create “Telegram Trigger”** (`Telegram Trigger` node)
|
||||
- Updates: `message`
|
||||
- Add Telegram credentials (Bot Token) under a credential like **Telegram account**.
|
||||
|
||||
2. **Create “Code in JavaScript”** (`Code` node)
|
||||
- Paste logic to extract:
|
||||
- `text` from `message.text` (default empty string)
|
||||
- `chat_id` from `message.chat.id` (default 0)
|
||||
- Connect: **Telegram Trigger → Code in JavaScript**
|
||||
|
||||
3. **Create “Get many products”** (`Shopify` node)
|
||||
- Resource: `Product`
|
||||
- Operation: `Get All`
|
||||
- Authentication: OAuth2
|
||||
- Connect Shopify OAuth2 credentials (store + scopes for reading products)
|
||||
- Connect: **Code in JavaScript → Get many products**
|
||||
|
||||
4. **Create “OpenAI Chat Model”** (LangChain OpenAI chat model node)
|
||||
- Model: `gpt-4o-mini`
|
||||
- Connect OpenAI credential (API key) under **OpenAi account**
|
||||
|
||||
5. **Create “Postgres Chat Memory”** (LangChain Postgres memory node)
|
||||
- Session ID type: `Custom Key`
|
||||
- Session key expression: `{{ $('Code in JavaScript').item.json.chat_id }}`
|
||||
- Context window length: `30`
|
||||
- Configure Postgres credential (**Postgres account 2**) to a reachable database.
|
||||
|
||||
6. **Create “AI Agent”** (LangChain Agent node)
|
||||
- Text input: `{{ $('Telegram Trigger').item.json.message.text }}`
|
||||
- (Optional improvement: use `{{ $('Code in JavaScript').item.json.text }}` for normalized input.)
|
||||
- Set the **System message** to the SOFÍA instructions, including:
|
||||
- Mandatory greeting protocol
|
||||
- Catalog precision and focus rules
|
||||
- Strict JSON output schema
|
||||
- Product list injection expression referencing **Get many products**:
|
||||
- `{{ $('Get many products').all().map(p => p.json.title + " (Precio: $" + p.json.variants[0].price + ")").join(', ') }}`
|
||||
- Enable output parsing (so the node returns structured output if supported).
|
||||
- Connect the special ports:
|
||||
- **OpenAI Chat Model (ai_languageModel) → AI Agent**
|
||||
- **Postgres Chat Memory (ai_memory) → AI Agent**
|
||||
- Connect main flow:
|
||||
- **Get many products → AI Agent**
|
||||
|
||||
7. **Create “Code in JavaScript 2”** (`Code` node)
|
||||
- Implement JSON extraction/parse from agent output (`items[0].json.output`)
|
||||
- Add fallback JSON object if parsing fails.
|
||||
- **Important:** If your n8n requires item arrays, return:
|
||||
- `return [{ json: parsed }];` and `return [{ json: fallback }];`
|
||||
- Connect: **AI Agent → Code in JavaScript 2**
|
||||
|
||||
8. **Create “Send a text message”** (`Telegram` node)
|
||||
- Operation: Send Message
|
||||
- Chat ID: `{{ $('Code in JavaScript').first().json.chat_id }}`
|
||||
- Text: `{{ $json.response }}`
|
||||
- Connect Telegram credential **Telegram account**
|
||||
- Connect: **Code in JavaScript 2 → Send a text message**
|
||||
|
||||
9. **Test**
|
||||
- Activate workflow
|
||||
- Send a Telegram text message to the bot
|
||||
- Verify:
|
||||
- First-time user gets the name question
|
||||
- Returning user gets personalized greeting
|
||||
- Product listing matches Shopify titles/prices
|
||||
|
||||
---
|
||||
|
||||
## 5. General Notes & Resources
|
||||
|
||||
| Note Content | Context or Link |
|
||||
|---|---|
|
||||
| This workflow turns a generic chatbot into an AI advisor that remembers users and responds using real business data. It combines user messages, persistent memory, and live product data to generate accurate and human-like responses. | Sticky note: “AI Advisor with Memory and Real-Time Context” |
|
||||
| How to set up: Import workflow, connect Telegram/OpenAI/Shopify/Postgres credentials, configure DB for chat memory, test via Telegram message. | Sticky note: “AI Advisor with Memory and Real-Time Context” |
|
||||
| Customization: Adjust system message tone/industry, replace Shopify with another data source, change memory window size. | Sticky note: “AI Advisor with Memory and Real-Time Context” |
|
||||
| ⚠️ Important: This workflow requires valid JSON output from the AI. If you modify the system prompt, make sure the response format is preserved to avoid parsing errors. | Sticky note: “⚠️ Important” |
|
||||
Reference in New Issue
Block a user