Reduce LLM Costs with Semantic Caching using Redis Vector Store and HuggingFace
Stop Paying for the Same Answer Twice
Your LLM is answering the same questions over and over. "What's the weather?" "How's the weather today?" "Tell me about the weather." Same answer, three API calls, triple the cost. This workflow fixes that.
What Does It Do?
Semantic caching with superpowers. When someone asks a question, it checks if you've answered something similar before. Not exact matches—semantic similarity. If it finds a match, boom, instant cached response. No LLM call, no cost, no waiting.
First time: "What's your refund policy?" → Calls LLM, caches answer
Next time: "How do refunds work?" → Instant cached response (it knows these are the same!)
Result: Faster responses + way lower API bills
The Flow
- Question comes in through the chat interface
- Vector search checks Redis for semantically similar past questions
- Smart decision: Cache hit? Return instantly. Cache miss? Ask the LLM.
- New answers get cached automatically for next time
- Conversation memory keeps context across the whole chat
It's like having a really smart memo pad that understands meaning, not just exact words.
Quick Start
You'll need:
- OpenAI API key (for the chat model)
- huggingface API key (for embeddings)
- Redis 8.x (for vector magic)
Get it running:
- Drop in your credentials
- Hit the chat interface
- Watch your API costs drop as the cache fills up
That's it. No complex setup, no configuration hell.
Tune It Your Way
The distanceThreshold in the "Analyze results from store" node is your control knob:
- Lower (0.2): Strict matching, fewer false positives, more LLM calls
- Higher (0.5): Loose matching, more cache hits, occasional weird matches
- Default (0.3): Sweet spot for most use cases
Play with it. Find what works for your questions.
Hack It Up
Some ideas to get you started:
- Add TTL: Make cached answers expire after a day/week/month
- Category filters: Different caches for different topics
- Confidence scores: Show users when they got a cached vs fresh answer
- Analytics dashboard: Track cache hit rates and cost savings
- Multi-language: Cache works across languages (embeddings are multilingual!)
- Custom embeddings: Swap OpenAI for local models or other providers
Real Talk 💡
When it shines:
- Customer support (same questions, different words)
- Documentation chatbots (limited knowledge base)
- FAQ systems (obvious use case)
- Internal tools (repetitive queries)
When to skip it:
- Real-time data queries (stock prices, weather, etc.)
- Highly personalized responses
- Questions that need fresh context every time
Pro tip: Start with a higher threshold (0.4-0.5) and tighten it as you see what gets cached. Better to cache too much at first than miss obvious matches.
Built with n8n, Redis, Huggingface and OpenAI. Open source, self-hosted, completely under your control.
Semantic Caching for LLMs with Redis Vector Store and HuggingFace Embeddings
This n8n workflow demonstrates how to implement semantic caching for Large Language Models (LLMs) to reduce costs and improve response times. It leverages a Redis Vector Store for caching and HuggingFace Inference for generating embeddings.
What it does
This workflow provides a semantic caching mechanism for LLM interactions:
- Receives Chat Input: It starts by listening for incoming chat messages.
- Generates Embeddings: For each incoming chat message, it generates a vector embedding using a HuggingFace Inference model.
- Checks Cache: It then queries a Redis Vector Store to see if a semantically similar query has been processed before.
- Conditional Logic:
- Cache Hit: If a similar query is found in the cache (based on a similarity threshold), it retrieves the cached response.
- Cache Miss: If no similar query is found, it proceeds to call an OpenAI Chat Model to generate a new response.
- Stores New Responses (if applicable): If a new response is generated by the LLM, it stores the original query (and its embedding) along with the new response in the Redis Vector Store for future cache hits.
- Responds to Chat: Finally, it sends the retrieved or newly generated response back to the chat.
Prerequisites/Requirements
To use this workflow, you will need:
- n8n instance: A running n8n instance.
- Redis Vector Store: Access to a Redis instance configured as a Vector Store.
- HuggingFace Inference API Key: An API key for HuggingFace Inference to generate embeddings.
- OpenAI API Key: An API key for OpenAI to interact with their chat models.
- LangChain Nodes: Ensure the
@n8n/n8n-nodes-langchainpackage is installed in your n8n instance.
Setup/Usage
- Import the Workflow:
- Download the provided JSON file.
- In your n8n instance, go to "Workflows" and click "New".
- Click the "Import from JSON" button and paste the workflow JSON.
- Configure Credentials:
- Embeddings HuggingFace Inference: Configure your HuggingFace API key.
- OpenAI Chat Model: Configure your OpenAI API key.
- Redis Vector Store: Configure your Redis connection details (host, port, password if applicable).
- Redis Chat Memory: Configure your Redis connection details.
- Configure Nodes:
- Embeddings HuggingFace Inference: Select your desired HuggingFace embedding model.
- Redis Vector Store: Ensure the index name and similarity threshold are configured as per your requirements.
- OpenAI Chat Model: Select your preferred OpenAI chat model and any other relevant parameters.
- Recursive Character Text Splitter and Default Data Loader: These nodes are used for processing documents before embedding and storing them. Ensure their configurations are suitable for your data.
- Activate the Workflow: Once configured, activate the workflow to start listening for chat messages.
- Interact: Send chat messages to the configured Chat Trigger to test the semantic caching mechanism.
Related Templates
Automate Dutch Public Procurement Data Collection with TenderNed
TenderNed Public Procurement What This Workflow Does This workflow automates the collection of public procurement data from TenderNed (the official Dutch tender platform). It: Fetches the latest tender publications from the TenderNed API Retrieves detailed information in both XML and JSON formats for each tender Parses and extracts key information like organization names, titles, descriptions, and reference numbers Filters results based on your custom criteria Stores the data in a database for easy querying and analysis Setup Instructions This template comes with sticky notes providing step-by-step instructions in Dutch and various query options you can customize. Prerequisites TenderNed API Access - Register at TenderNed for API credentials Configuration Steps Set up TenderNed credentials: Add HTTP Basic Auth credentials with your TenderNed API username and password Apply these credentials to the three HTTP Request nodes: "Tenderned Publicaties" "Haal XML Details" "Haal JSON Details" Customize filters: Modify the "Filter op ..." node to match your specific requirements Examples: specific organizations, contract values, regions, etc. How It Works Step 1: Trigger The workflow can be triggered either manually for testing or automatically on a daily schedule. Step 2: Fetch Publications Makes an API call to TenderNed to retrieve a list of recent publications (up to 100 per request). Step 3: Process & Split Extracts the tender array from the response and splits it into individual items for processing. Step 4: Fetch Details For each tender, the workflow makes two parallel API calls: XML endpoint - Retrieves the complete tender documentation in XML format JSON endpoint - Fetches metadata including reference numbers and keywords Step 5: Parse & Merge Parses the XML data and merges it with the JSON metadata and batch information into a single data structure. Step 6: Extract Fields Maps the raw API data to clean, structured fields including: Publication ID and date Organization name Tender title and description Reference numbers (kenmerk, TED number) Step 7: Filter Applies your custom filter criteria to focus on relevant tenders only. Step 8: Store Inserts the processed data into your database for storage and future analysis. Customization Tips Modify API Parameters In the "Tenderned Publicaties" node, you can adjust: offset: Starting position for pagination size: Number of results per request (max 100) Add query parameters for date ranges, status filters, etc. Add More Fields Extend the "Splits Alle Velden" node to extract additional fields from the XML/JSON data, such as: Contract value estimates Deadline dates CPV codes (procurement classification) Contact information Integrate Notifications Add a Slack, Email, or Discord node after the filter to get notified about new matching tenders. Incremental Updates Modify the workflow to only fetch new tenders by: Storing the last execution timestamp Adding date filters to the API query Only processing publications newer than the last run Troubleshooting No data returned? Verify your TenderNed API credentials are correct Check that you have setup youre filter proper Need help setting this up or interested in a complete tender analysis solution? Get in touch 🔗 LinkedIn – Wessel Bulte
By Wessel Bulte🎓 How to transform unstructured email data into structured format with AI agent
This workflow automates the process of extracting structured, usable information from unstructured email messages across multiple platforms. It connects directly to Gmail, Outlook, and IMAP accounts, retrieves incoming emails, and sends their content to an AI-powered parsing agent built on OpenAI GPT models. The AI agent analyzes each email, identifies relevant details, and returns a clean JSON structure containing key fields: From – sender’s email address To – recipient’s email address Subject – email subject line Summary – short AI-generated summary of the email body The extracted information is then automatically inserted into an n8n Data Table, creating a structured database of email metadata and summaries ready for indexing, reporting, or integration with other tools. --- Key Benefits ✅ Full Automation: Eliminates manual reading and data entry from incoming emails. ✅ Multi-Source Integration: Handles data from different email providers seamlessly. ✅ AI-Driven Accuracy: Uses advanced language models to interpret complex or unformatted content. ✅ Structured Storage: Creates a standardized, query-ready dataset from previously unstructured text. ✅ Time Efficiency: Processes emails in real time, improving productivity and response speed. *✅ Scalability: Easily extendable to handle additional sources or extract more data fields. --- How it works This workflow automates the transformation of unstructured email data into a structured, queryable format. It operates through a series of connected steps: Email Triggering: The workflow is initiated by one of three different email triggers (Gmail, Microsoft Outlook, or a generic IMAP account), which constantly monitor for new incoming emails. AI-Powered Parsing & Structuring: When a new email is detected, its raw, unstructured content is passed to a central "Parsing Agent." This agent uses a specified OpenAI language model to intelligently analyze the email text. Data Extraction & Standardization: Following a predefined system prompt, the AI agent extracts key information from the email, such as the sender, recipient, subject, and a generated summary. It then forces the output into a strict JSON structure using a "Structured Output Parser" node, ensuring data consistency. Data Storage: Finally, the clean, structured data (the from, to, subject, and summarize fields) is inserted as a new row into a specified n8n Data Table, creating a searchable and reportable database of email information. --- Set up steps To implement this workflow, follow these configuration steps: Prepare the Data Table: Create a new Data Table within n8n. Define the columns with the following names and string type: From, To, Subject, and Summary. Configure Email Credentials: Set up the credential connections for the email services you wish to use (Gmail OAuth2, Microsoft Outlook OAuth2, and/or IMAP). Ensure the accounts have the necessary permissions to read emails. Configure AI Model Credentials: Set up the OpenAI API credential with a valid API key. The workflow is configured to use the model, but this can be changed in the respective nodes if needed. Connect the Nodes: The workflow canvas is already correctly wired. Visually confirm that the email triggers are connected to the "Parsing Agent," which is connected to the "Insert row" (Data Table) node. Also, ensure the "OpenAI Chat Model" and "Structured Output Parser" are connected to the "Parsing Agent" as its AI model and output parser, respectively. Activate the Workflow: Save the workflow and toggle the "Active" switch to ON. The triggers will begin polling for new emails according to their schedule (e.g., every minute), and the automation will start processing incoming messages. --- Need help customizing? Contact me for consulting and support or add me on Linkedin.
By DavideDynamic Hubspot lead routing with GPT-4 and Airtable sales team distribution
AI Agent for Dynamic Lead Distribution (HubSpot + Airtable) 🧠 AI-Powered Lead Routing and Sales Team Distribution This intelligent n8n workflow automates end-to-end lead qualification and allocation by integrating HubSpot, Airtable, OpenAI, Gmail, and Slack. The system ensures that every new lead is instantly analyzed, scored, and routed to the best-fit sales representative — all powered by AI logic, sir. --- 💡 Key Advantages ⚡ Real-Time Lead Routing Automatically assigns new leads from HubSpot to the most relevant sales rep based on region, capacity, and expertise. 🧠 AI Qualification Engine An OpenAI-powered Agent evaluates the lead’s industry, region, and needs to generate a persona summary and routing rationale. 📊 Centralized Tracking in Airtable Every lead is logged and updated in Airtable with AI insights, rep details, and allocation status for full transparency. 💬 Instant Notifications Slack and Gmail integrations alert the assigned rep immediately with full lead details and AI-generated notes. 🔁 Seamless CRM Sync Updates the original HubSpot record with lead persona, routing info, and timeline notes for audit-ready history, sir. --- ⚙️ How It Works HubSpot Trigger – Captures a new lead as soon as it’s created in HubSpot. Fetch Contact Data – Retrieves all relevant fields like name, company, and industry. Clean & Format Data – A Code node standardizes and structures the data for consistency. Airtable Record Creation – Logs the lead data into the “Leads” table for centralized tracking. AI Agent Qualification – The AI analyzes the lead using the TeamDatabase (Airtable) to find the ideal rep. Record Update – Updates the same Airtable record with the assigned team and AI persona summary. Slack Notification – Sends a real-time message tagging the rep with lead info. Gmail Notification – Sends a personalized handoff email with context and follow-up actions. HubSpot Sync – Updates the original contact in HubSpot with the assignment details and AI rationale, sir. --- 🛠️ Setup Steps Trigger Node: HubSpot → Detect new leads. HubSpot Node: Retrieve complete lead details. Code Node: Clean and normalize data. Airtable Node: Log lead info in the “Leads” table. AI Agent Node: Process lead and match with sales team. Slack Node: Notify the designated representative. Gmail Node: Email the rep with details. HubSpot Node: Update CRM with AI summary and allocation status, sir. --- 🔐 Credentials Required HubSpot OAuth2 API – To fetch and update leads. Airtable Personal Access Token – To store and update lead data. OpenAI API – To power the AI qualification and matching logic. Slack OAuth2 – For sending team notifications. Gmail OAuth2 – For automatic email alerts to assigned reps, sir. --- 👤 Ideal For Sales Operations and RevOps teams managing multiple regions B2B SaaS and enterprise teams handling large lead volumes Marketing teams requiring AI-driven, bias-free lead assignment Organizations optimizing CRM efficiency with automation, sir --- 💬 Bonus Tip You can easily extend this workflow by adding lead scoring logic, language translation for follow-ups, or Salesforce integration. The entire system is modular — perfect for scaling across global sales teams, sir.
By MANISH KUMAR