In the world of technical support, tickets describing the same issue can be written in many different ways. One user writes 'I can't send international SMS', another reports 'Texts to Germany fail', and a third says 'Messages to +49 numbers are stuck'. A traditional keyword-based search engine will not connect these variants, forcing support teams to manually review hundreds of tickets or rely on rigid rules. Semantic search solves this challenge by understanding the meaning behind phrases, not just exact terms. In this article we explore how to implement a semantic search system for support tickets using Python, AI embeddings, and a lightweight architecture, with practical examples you can adapt to your business.
The core idea is to convert each ticket —combining its subject and body— into a numerical vector that captures its meaning. Then, when a query arrives, we transform it into another vector and calculate cosine similarity to retrieve the most relevant tickets. This approach, popularized by models like those from OpenAI, is now accessible to any developer thanks to inference APIs such as Telnyx. At Q2BSTUDIO, as a company specialized in custom software and artificial intelligence, we know that integrating these capabilities into support platforms can drastically reduce resolution times and improve user experience. For example, an agent searching for 'audio problems on calls' will automatically find tickets about 'voice quality' or 'echo on calls' without manual tagging.
To build this solution in Python, we use Flask as the web framework and the Telnyx embeddings API (model thenlper/gte-large). The flow is simple: first, we index a set of support tickets stored in a JSON file. Each ticket has fields like subject, body, category, and priority. We concatenate subject and body, send that string to the /v2/ai/openai/embeddings endpoint, and store the resulting vector in a NumPy matrix in memory. When a user performs a search, we repeat the process with the query, compute cosine similarity against all indexed vectors, and return the top_k tickets ordered by relevance. This system does not require a full vector database; it is perfect for prototypes or small teams wanting to try semantic search without heavy infrastructure investment.
The API routes include POST /index to build the index from scratch, POST /search to search, POST /tickets to add a new ticket (which is embedded and appended to the index on the fly), GET /tickets/<id> to fetch a ticket by ID, GET /stats to view index statistics, and GET /health to check service status. The implementation is lightweight and can run locally with just a few dependencies and a Telnyx API key. For production environments, Q2BSTUDIO recommends migrating vector storage to systems like pgvector, Qdrant, or Pinecone, and combining semantic search with filters for category, priority, or date. Additionally, cybersecurity of the API and data is crucial: ensure keys are protected and queries are properly sanitized.
A direct use case is duplicate ticket detection. If a support team receives multiple reports about the same incident, semantic search can automatically group them, even if users use different language. It is also useful for intelligent routing: a ticket about 'number transfer failure' can be directed to the portability team, while one about 'audio quality' goes to the voice team. Companies handling large volumes of tickets can benefit from integrating this logic into their BI/Power BI tools, generating dashboards that show semantic trends in reported issues. At Q2BSTUDIO we develop custom solutions that connect these capabilities with cloud platforms like AWS or Azure, ensuring scalability and high availability.
Beyond support, semantic search applies to conversational AI, virtual assistants, and recommendation systems. For example, an AI agent can use this same mechanism to suggest automatic replies based on similar past tickets. The combination of large language models (LLMs) with embedding-based semantic search is a growing trend known as RAG (Retrieval-Augmented Generation). At Q2BSTUDIO we have implemented these patterns for clients needing AI agents capable of understanding customer context without relying on fixed scripts. Our engineering team also integrates cybersecurity solutions to protect both embeddings and sensitive data that tickets may contain.
If you decide to take this prototype to production, consider that the in-memory matrix is not persistent. A server restart would wipe the index. To avoid this, you can store vectors in a vector database or even in a cloud object store. Additionally, the embedding model can be updated: the Telnyx API supports multiple models, and you can experiment with others like text-embedding-ada-002 if you need higher accuracy. Latency is also a factor: with hundreds of thousands of tickets, cosine similarity computation can become slow. That is where optimizations like approximate nearest neighbor (HNSW) or semantic clustering come into play.
To get started, clone the example repository from Telnyx, set up your API key, and run the application. Then try searches like 'my API key stopped working after I changed it' or 'phone number transfer is taking too long'. You will see how the system returns relevant tickets even if the words do not match exactly. At Q2BSTUDIO we help companies make this technological leap, from conceptualization to deployment in cloud environments, always with a focus on software quality and data security. Semantic search not only improves support efficiency but also lays the groundwork for future innovations in artificial intelligence applied to customer service.
In summary, this architecture shows that you do not need complex infrastructure to start benefiting from meaning-based search. A simple Python script, an embeddings API, and a bit of creativity can transform how your team handles support tickets. If you are looking to implement robust, scalable solutions tailored to your business, contact Q2BSTUDIO. Our expertise in custom software, cloud AWS/Azure, cybersecurity, and BI/Power BI allows us to offer comprehensive support at every project stage.
