Practical notes from the engineers who build agents, copilots and AI products.
The retrieval engineering is the same; the surrounding work is not. Startups buy one corpus, one or two connectors and a shipped feature in six to eight weeks. Enterprises buy permission-aware retrieval across many systems, an audit trail and a change-management plan.
RAG projects fail for five reasons: the corpus cannot answer the questions, retrieval quality was never measured, permissions were an afterthought, nobody owns the content after launch, and the task was never a retrieval problem. Four of the five are visible before the first sprint, if somebody checks.
No retrieval system is DPDP-compliant as a category; an implementation is compliant or it is not. Under India's Digital Personal Data Protection Act, 2023 a RAG pipeline is processing on behalf of a data fiduciary, so retention, erasure and breach duties apply to the index, not just the source.
Measure four layers: retrieval, generation, product and business. Retrieval tells you whether the right passage was found, generation whether the answer stayed inside it, product whether anyone uses the system, and business whether a number your CFO tracks moved. A programme failing at one layer is failing.
Implement a text to SQL solution in five stages: inventory the questions people actually ask, build a semantic layer that fixes definitions, retrieve the relevant schema instead of prompting the whole warehouse, generate and validate read-only SQL, then gate every release on an eval set of known-correct queries.
Text to SQL solution projects fail for five reasons, and none of them is model quality: metrics with no agreed definition, a demo schema that is not the production schema, permissions enforced in the prompt, no evaluation suite to catch regressions, and a launch to everyone at once.
Startups and enterprises build the same text to SQL solution differently in four places: scope, governance, security and integration depth. A startup ships forty questions over one schema in eight weeks from $12,500 or ₹8 lakh. An enterprise spends most of its budget on governance and takes twelve to sixteen weeks.
Measure a text to SQL solution on three layers: execution accuracy against a golden question set, trust signals such as refusal rate and repeat usage, and the business outcome the system was funded for. Report one headline number, verified answer rate, with the other layers underneath it.
A text to SQL solution can be DPDP-compliant, but only by design. The Act cares about four things your architecture decides: where processing happens, which rows the asker may see, what the logs keep, and whether the purpose was stated. Get those right and the risk is no worse than your existing BI tool.