AINPM.in / Glossary
Glossary
AI-native product management glossary
33 terms a product manager meets when building AI products in India, each defined in one or two sentences.
- Abstention
- Agentic AI
- AI-native product
- AI-native product management
- BM25
- Chunking
- Citation precision
- Citation recall
- Consent manager
- Cross-encoder reranker
- Digital public infrastructure (DPI)
- DPDP Act
- Eval (evaluation)
- Faithfulness
- FREE-AI
- Global capability centre (GCC)
- Golden set
- Hybrid retrieval
- Inference cost
- Key wrapping
- Open Badges 3.0
- Passkey
- Proof of skill
- Quality gate
- Recall@k
- Retrieval-augmented generation (RAG)
- Skills framework
- Skills-based hiring
- Supersession
- Synthetically generated information (SGI)
- Verifiable credential
- WebAuthn PRF extension
- Zero-knowledge storage
- Abstention
- When an AI product deliberately declines to answer because the evidence is missing or the question is out of scope. Abstention accuracy measures how often it refuses, or answers, correctly.
- Agentic AI
- AI systems that plan and take multi-step actions with tools, such as searching, calling APIs or writing files, to complete a goal, usually with human approval at defined points.
- AI-native product
- A product whose core value is delivered by an AI model, so it would not exist without the model. It contrasts with an AI-enabled product, where AI is added to an existing workflow.
- AI-native product management
- The practice of defining, building and improving AI-native products: specifying behaviour with evaluation sets, measuring quality as rates, managing inference cost per task and designing trust controls from the start.
- BM25
- A keyword-based ranking method that scores documents by how often query terms appear, adjusted for document length. It is often combined with vector search in hybrid retrieval.
- Chunking
- Splitting documents into smaller passages before indexing so that retrieval returns focused evidence. Section-aware chunking follows the document’s own structure.
- Citation precision
- Of the sources an answer cites, the share that are actually relevant. Low citation precision means extra, unnecessary citations.
- Citation recall
- Of the sources an answer should cite, the share it does cite. Low citation recall means the governing source is being missed.
- Consent manager
- Under India’s DPDP framework, a registered intermediary that lets people give, review and withdraw consent across services. The framework starts on 13 November 2026.
- Cross-encoder reranker
- A model that reads a query and a candidate passage together and scores their relevance. It is slower than vector search but more accurate, so it is used to reorder the top results.
- Digital public infrastructure (DPI)
- India’s shared digital rails, including UPI for payments, Aadhaar for identity, DigiLocker for documents, Account Aggregator for consented financial data and ONDC for commerce.
- DPDP Act
- India’s Digital Personal Data Protection Act, 2023. Its Rules were notified on 14 November 2025 and apply in phases, with full obligations from 13 May 2027.
- Eval (evaluation)
- A repeatable test that scores an AI product against a fixed set of examples with known good answers, reporting quality as rates so changes can be compared.
- Faithfulness
- Whether an AI answer is supported by the sources it was given. A faithfulness check flags claims or citations that the retrieved context does not back.
- FREE-AI
- The Reserve Bank of India’s Framework for Responsible and Ethical Enablement of AI, published in August 2025, with seven guiding principles for AI in regulated finance.
- Global capability centre (GCC)
- An offshore unit of a multinational company that performs technology, product and business work. India had 2,117 GCCs employing 2.36 million people in FY2026.
- Golden set
- A reviewed collection of test questions with expected answers or sources, used as the reference for evaluating an AI product. It acts as the product specification.
- Hybrid retrieval
- Combining keyword search such as BM25 with vector search, then fusing the two ranked lists. It handles both exact terms and paraphrased questions.
- Inference cost
- The cost of running a model to produce one output, driven by model choice, input and output tokens and caching. AI-native PMs track it per task.
- Key wrapping
- Encrypting one cryptographic key with another. A vault key wrapped under a password- or passkey-derived key can be stored safely, because the stored ciphertext is useless without the wrapping key.
- Open Badges 3.0
- A 1EdTech standard for digital badges that follows the W3C Verifiable Credentials model, so a badge carries signed, machine-readable proof of who issued it and which skills it covers.
- Passkey
- A WebAuthn credential that replaces a password with a key pair held by the device or a password manager and unlocked with biometrics or a PIN. Passkeys are often synced across a user’s devices.
- Proof of skill
- Public evidence that someone can do what they claim, such as a live demo, an evaluation log, open code or a verifiable credential, linked to a specific, plainly worded skill.
- Quality gate
- A minimum score for each evaluation metric that a change must clear before release. Gates should be re-derived when the model, data or test set changes.
- Recall@k
- The share of questions for which a relevant document appears in the top k retrieved results. Recall@10 of 0.943 means the right document is in the top ten 94.3% of the time.
- Retrieval-augmented generation (RAG)
- A pattern where the system first retrieves relevant passages from a trusted corpus and then asks a language model to answer using only those passages, usually with citations.
- Skills framework
- A shared, structured list of skills and work activities, such as O*NET, ESCO or NACE Career Readiness Competencies, used to describe skills in language employers and hiring systems recognise.
- Skills-based hiring
- Hiring that judges candidates on demonstrated skills rather than degrees or job titles alone, usually by asking for evidence, work samples or skill-tagged credentials.
- Supersession
- When a newer regulation or circular replaces an older one. Supersession-aware systems flag the older document and point to its replacement.
- Synthetically generated information (SGI)
- The term used in India’s IT Rules for AI-generated or AI-altered content. From 20 February 2026 it must carry prominent labels.
- Verifiable credential
- A digital certificate signed by its issuer, following the W3C Verifiable Credentials standard, so anyone can confirm it is genuine without contacting the issuer.
- WebAuthn PRF extension
- A WebAuthn extension that asks a passkey authenticator for a high-entropy secret, computed in hardware after user verification. Because it is never stored, it can serve as the root of client-side encryption.
- Zero-knowledge storage
- A storage design in which the provider, or anyone holding a copy of the stored data, cannot read it. Only the user, with a password or passkey, can derive the key.
See these terms in use in What is AI-native product management? and the SEBI Circular RAG case study.