Embeddings and RAG · Indu Learning Platform

The problem RAG solves

Your specs, IMG notes and tickets are unknown to the model and far too big to paste into every prompt. Retrieval-augmented generation (RAG) first retrieves the few most relevant passages, then generates an answer from only those, with citations. The model's job shifts from remembering to reading, which it does well.

Embeddings: meaning as numbers

An embedding is a vector of numbers that an embedding model produces from text. Similar meanings get vectors pointing in similar directions: 'supplier blocked for payment' lands near 'vendor payment block'. Cosine similarity (-1 to 1) measures that closeness; near 1 means same topic. It measures meaning, not correctness.

Illustrative similarity to 'supplier blocked for payment'. Real values depend on the model.

Keyword versus semantic search

SAP content is full of exact codes, so production systems often combine both (hybrid search).

Hybrid search is usually the safest choice for SAP content.

Offline: indexing your documents

Done once, refreshed whenever documents change.

Online: following one question

Step through · 1 of 7

A data steward asks: 'Which CR type do we use to block a supplier for payment in company code 1000?'

One RAG query, from question to cited answer.

Try it: build the pipeline

Try it · Put in order

Put the full RAG pipeline in order, from raw documents to a cited answer.

1Embed each chunk
2Store vectors with metadata
3Split specs and tickets into chunks
4Embed the user's question with the same model
5Retrieve the top-k most similar chunks
6Put the chunks and question into the prompt
7Generate an answer that cites its sources

Chunking and metadata decide quality

Chunks must be small enough to retrieve precisely, big enough to make sense alone. Split along the document's own structure, with a small overlap so rules are not cut in half. Tag each chunk with metadata (doc ID, section, object, release, status) to filter and cite.

Chunk along the structure the authors already used.

When RAG answers badly

Check the retrieved chunks first. If the right passage is missing, it is a retrieval problem: fix chunking, search or metadata. If it was retrieved and the answer is still wrong, it is a generation problem: tighten the prompt, require citations, or use a stronger model.

Python

import numpy as np

def cosine(a, b):
    a, b = np.array(a), np.array(b)
    return float(a @ b / (np.linalg.norm(a) * np.linalg.norm(b)))

# embed() = your platform's embedding API; chunks are embedded once, offline
chunks = [
    {"id": "FS-BP-017 4.2", "text": "Payment block for suppliers is set via CR type ZSUP_BLK ..."},
    {"id": "TCK-2231",      "text": "Supplier posting block not replicated to ECC via DRF ..."},
]
for c in chunks:
    c["vec"] = embed(c["text"])
q = embed("How do we block a supplier for payment?")
top = sorted(chunks, key=lambda c: cosine(q, c["vec"]), reverse=True)[:2]

Real systems use a vector database, but the logic is the same.

Try it: retrieval or generation?

Try it · Sort

A RAG assistant gave a bad answer. Where is the problem?

Retrieval problem

Generation problem

Click an item, then a bucket (or drag it).

Try it: the outdated answer

Try it · What would you do?

Your spec assistant confidently quotes the 2023 rule for supplier group ZINT payment terms. The 2025 spec changed it, and both versions are in the index. What do you do?

Remember

RAG = retrieve relevant passages from your documents, then answer from them with citations.
Embeddings capture meaning; chunking, metadata and hybrid search decide retrieval quality.
Debug retrieval first, generation second, and filter by access rights.

Check yourself

What is the main benefit of RAG on project documents?

A RAG assistant answers wrongly. The right spec passage was not among the retrieved chunks. Where do you fix it?

Why might pure embedding search struggle with 'CREMAS05 segment E1LFA1M'?

Two chunks have cosine similarity 0.92 with a question. What does that indicate?