AI-first Series – Part 2
In the previous blog, our data engineering services team outlined the foundational steps to becoming an AI-first enterprise. We highlighted how a majority of organizations, despite crafting grand AI roadmaps, remain stuck in AI-last mode. The real transition to becoming an AI-first enterprise begins with data labelling. From manual to AI-assisted labelling – the core objective of the data engineering services team is to ensure that you have tags associated with your data.
Once the AI-assisted labelling is complete, you have effectively created a list of smart records. Through these smart records, the AI model can perform tasks such as searching, filtering and categorizing specific to these individual records. In other words, the AI model can understand the individual records, but may still lack the ability to establish connections between these records.
For example, there is a support ticket that says “I cannot login post upgrade”. The AI model can comprehend the user sentiment through tags such as “Frustrated” or “Login issue”. It also understands it is just a ticket. However, it has little awareness of the broader context or severity of the problem: whether the user has experienced the same problem previously or if other users have faced the same issue.
This is Where Logical Data Modelling Comes into Play!
Logical Data Modelling helps the AI model to correlate different records and establish connections between them. Also, in data modelling, your users, products, and tickets will become separate entities with their own records. In the above case, the AI models can tap into all “support tickets “ and “users” to explore if the problem persists across users (See Fig, 1). It can also probe past login issues tied to the same user/product combo (See Fig. 2).
SELECT * FROM support_tickets
WHERE product_id = ‘P-AcmeApp’ AND version = ‘v4.2’
AND tag = ‘login issue’
SELECT * FROM support_tickets
WHERE user_id = ‘U001’ AND tag = ‘login issue’
While with tags, your records are searchable by the AI model, with data modelling, your AI model can understand relationships between entities and thus achieves contextual awareness of your data.
| Remember, as in many real world cases where companies struggle to move beyond tagging, partnering with expert data engineering services will help you not just create smart records but build contextual intelligence with deeper data modelling. |
Data Modelling is Done, Thanks to Data Engineering Services. Are We Good with Unveiling our First AI-Assistant?
Well, not quite! Your AI can only establish connections between structured data. But what about unstructured data? Imagine a user reaching out through your website chat and saying “I can’t use the product all of a sudden.” This user is experiencing the same problem after the upgrade, but instead of raising a ticket, they had communicated the problem through the website chat. How will you train your AI model to recognize this message as equivalent to a ticket?
Again, alluding to our past data engineering services engagements, we have seen that this stage is where most companies tend to stumble. This is corroborated by a Forbes report which highlights that 90% of unstructured data remains untapped. The secret to making sense of your unstructured data lies in RAG!
How RAG Breathes Life into your AI-Assistant
Retrieval Augment Generation (RAG) employs two key components to help the AI model generate a relevant, grounded answer.
- The first component is the Embedding, which converts unstructured data (FAQs, documents, chat) into a dense numerical representation called the vectors. Words and phrases with similar meanings/concepts will have similar vectors. They will be positioned close together in the A vector space. For instance, “forgot login” and “reset password” will appear close together in the vector space, thus helping the AI model understand the words are semantically related.
- So, when a user enters saying “I can’t log in all of a sudden”, RAG first embeds the query, i.e., it converts the text into a vector. Then it scans through embeddings to find the most similar vector. These embeddings are stored in a specialized database called Vector DB that indexes millions of embeddings so the AI model can search and arrive at a list of relevant text chunks.
- These text chunks are then injected into the AI model along with the original user prompt to generate a contextually accurate answer.
Here’s How RAG Works Step by Step:
1 User Input →“I can’t log in all of a sudden”
2 Embed the Query → Turn user input into a dense vector using an embedding model. (example vectors – 0.12, -0.44, 0.87)
3 Search in Vector DB → Find semantically similar vectors. The vector DB contains precomputed vectors of past resolved tickets, FAQ entries on login issues, and known issue logs.
4 Retrieve Top Matches → Pull the top, relevant chunks.
Examples of retrieved chunks: “Login errors increased after v4.2 update; clearing cache resolved issue.” “Password reset link failed in older versions.”
5 Inject into the LLM Prompt → Combines the user input + retrieved chunks to generate an augmented prompt.
Examples: User query – I can’t log in all of a sudden + Retrieved chunks: “Clearing cache fixed login issues in v4.2”, “Version 4.2.1 patch notes: fixed login bug.”
6 LLM Generates Output → The AI model uses the augmented prompt to generate an intelligent response.
Example: “I’m sorry you’re having trouble logging in. This could be related to a known issue in version
4.2. Please try clearing your cache or check if you’re using SSO. You may also want to update to v4.2.1, which fixes this problem.”
Voila! You have created your first assistant that can help your users with contextually accurate answers in real-time.
Next, how do we set out to create Agents?
What does it take to build Agents, the true powerhouse of Your AI-first Enterprise?
Stay tuned for Part 3.
