How to Cache Chatbot Responses in n8n and Reduce Redundant Vector DB Queries?

Below is one approach to implement caching in your chatbot workflow:

---

1. Create a Cache Table

Use your PostgreSQL database to store queries and their responses. For example, run a SQL command (outside of n8n) to create a table:

SQLCopied
CREATE TABLE query_cache (
  query_hash TEXT PRIMARY KEY,
  query_text TEXT,
  cached_response JSONB,
  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

Tip: Storing a hash (such as an MD5 or SHA256 of the query) can speed up lookups and avoid storing redundant text.

---

2. Update Your Workflow

Modify your workflow to include these additional steps:

a. Check Cache

• Postgres Select node:
  Query the cache table using a query parameter (your query hash). For example:

  SQLCopied
  SELECT cached_response FROM query_cache WHERE query_hash = $1;

  In the Query Parameters field, pass an expression that computes the hash from the incoming query (you can use a Function node beforehand to generate this hash).

• If cache found:
  Use a conditional (If) node to check if the Select node returned any data.

  • If yes, output the cached response as the final result.
  • If not, continue to the next step.

b. Execute Semantic Search

• Continue with your existing vector database query (using your semantic search node) to retrieve the best-matching answer.

c. Cache Response

• Postgres Insert or Update node:
  Once you have the new response, insert it into the cache table. You might use a query like:

  SQLCopied
  INSERT INTO query_cache (query_hash, query_text, cached_response)
  VALUES ($1, $2, $3)
  ON CONFLICT (query_hash)
  DO UPDATE SET cached_response = EXCLUDED.cached_response, created_at = CURRENT_TIMESTAMP;

  Again, use Query Parameters to pass the hash, the query text, and the response.

• Finally, return the response to the chat.

---

3. Helpful Tips

• Compute the Query Hash:
  Use a Function node to compute a hash for the query text. For example, in JavaScript you could use a simple hash function or a crypto library if available.

• Managing Cache Lifetime:
  Depending on your use case, you might want to expire old cache entries. You can periodically run a cleanup query from within n8n or your database.

• Workflow Branching:
  Ensure the flow splits so that if the cache is hit, you skip the semantic search, saving on both database and token usage.

---

By incorporating these steps, your workflow will first attempt to retrieve a cached response. Only if no cached entry exists will it perform the vector search and then store the result for subsequent queries. This should help minimize redundant queries and reduce overall processing costs.