AI & personal dataUpdated 22 November 202611 min read

Handling DPDP Erasure Requests in LLMs and Vector Databases

How to honour a DPDP erasure request across vector databases, prompt logs, fine-tuning sets, eval sets and model weights — while respecting Rule 8(3)'s one-year retention minimum.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Statutory basis
DPDP Act ss. 8(7), 12 and 13
Procedural rules in force
13 May 2027 (Rules 8, 9, 14)
Minimum log retention
One year from processing (Rule 8(3))
Grievance response ceiling
Not more than 90 days (Rule 14(3))
Processor obligation
Fiduciary must cause processors to erase (s. 8(7))
Max penalty, other provisions
Up to ₹50 crore per instance

Quick answer

Under sections 8(7) and 12 of the DPDP Act, a Data Fiduciary must erase personal data when consent is withdrawn or the purpose is served, and make its Data Processors do the same. For AI systems that means deleting vectors, chunks, caches, prompt logs and training rows by subject ID, while Rule 8(3) still requires processing logs to be kept for one year before final erasure, from 13 May 2027.

  • Erasure duties come from sections 8(7) and 12 of the Act; the Rules add procedure through Rules 8, 9 and 14 from 13 May 2027.
  • In a RAG stack, erasure means deleting source chunks, embeddings, caches and index entries, not just the original document.
  • Rule 8(3) requires personal data, traffic data and processing logs to be retained for at least one year from processing, then erased.
  • Fine-tuned weights cannot be surgically edited today, so the practical control is keeping identifiable personal data out of training sets.
  • Every response to an erasure request must carry the Rule 9 contact details and fit inside a published grievance window of at most 90 days.

Who this is for: ML and platform engineers, CTOs of AI products, Data Protection Officers, Teams running RAG or fine-tuning pipelines, Privacy and compliance leads

What does a DPDP erasure request require from an AI system?

A Data Principal can ask a Data Fiduciary to erase her personal data under section 12 of the DPDP Act. Separately, section 8(7) requires erasure once consent is withdrawn or the specified purpose is no longer served, unless retention is required by law, and requires the fiduciary to cause its Data Processors to erase too.

For a conventional app, erasure is a row delete plus backups. For an AI product, the same person's data is usually copied into five or six derived stores:

  • Source corpora — tickets, documents, CRM notes that feed retrieval.
  • Vector databases — chunks and their embeddings, plus metadata.
  • Caches — semantic caches, response caches, conversation memory.
  • Prompt and trace logs — inputs, retrieved context, outputs, tool calls.
  • Training artefacts — fine-tuning datasets, preference data, eval sets.
  • Model weights — anything learned during fine-tuning.

An erasure request is only honoured when every derived copy is accounted for. The Act does not carve out embeddings or logs: if a vector can be traced back to an identifiable individual through its metadata or its source chunk, treat it as personal data.

The procedural layer — publishing how to make requests (Rule 14(1)), the 90-day grievance ceiling (Rule 14(3)) and the contact person in every response (Rule 9) — applies from 13 May 2027.

How does Rule 8(3)'s one-year retention interact with erasure?

Rule 8(3) says that, without prejudice to the Third Schedule erasure periods, a Data Fiduciary shall retain personal data, associated traffic data and other logs of the processing for a minimum of one year from the date of processing, for the purposes in the Seventh Schedule, and then cause them to be erased unless another law requires longer retention.

The Rule's own illustration is direct: an e-book platform must keep order details and processing logs for at least one year from the transaction "even if X deletes her account".

A workable reading for AI systems is a two-stage erasure:

  1. Operational erasure (immediately): remove the person's data from every store that serves live features — retrieval indexes, personalisation, memory, caches, and future training runs. Processing for the original purpose stops.
  2. Retention hold (up to one year from the processing date): move the minimum records required by Rule 8(3) into a restricted, encrypted archive that no product feature can read. Access is limited to security and legal roles, and every access is logged.
  3. Final erasure: a scheduled job deletes the held records when the one-year clock (per record) expires, unless another law requires longer.

Document this design in your erasure policy and tell the Data Principal, in plain language, what is removed now and what is held and why. This is an interpretation; involve counsel for edge cases such as litigation holds.

Erasing personal data from vector databases and RAG indexes

Vector stores make erasure hard only when ingestion did not plan for it. Build the delete path at ingestion time:

  • Tag every chunk with subject identifiers. Store subject_ids, source_doc_id, tenant_id and ingested_at as filterable metadata. Without subject tags you cannot find a person's vectors except by re-scanning text.
  • Delete by filter, then verify. Issue a metadata-filtered delete, then run a verification query for the same filter and for exact-match text search on name, email and phone.
  • Handle soft deletes. Many approximate-nearest-neighbour indexes mark deletions as tombstones and physically remove them only on compaction or segment merge. Schedule compaction so deleted vectors are not recoverable indefinitely.
  • Re-chunk mixed documents. If a chunk mentions several people, redact the requester's data from the source document, re-chunk and re-embed rather than deleting the whole chunk.
  • Purge caches. Semantic caches keyed on embeddings can replay an answer containing erased data. Invalidate entries whose stored context references the erased source_doc_id.
  • Snapshots and replicas. Index snapshots used for rollback must follow the same deletion schedule as backups.
StoreErasure mechanismVerification
Vector DBFiltered delete + compactionFiltered query returns zero
Source corpusDelete or redact, re-embedText search on identifiers
Semantic cacheInvalidate by doc IDCache key scan
Conversation memoryDelete by user IDMemory API returns empty

Prompt logs, fine-tuning sets and eval sets

Prompt and trace logs

Prompt logs are where most unplanned personal data lives. Mask identifiers before logs are written (see masking and tokenisation for AI pipelines) and key every trace by an internal subject ID. On erasure, purge or move traces into the restricted retention archive described above. The design trade-offs are covered in LLM logging and observability.

Fine-tuning datasets

Keep a lineage table mapping each training example to its source record and subject ID. On erasure, remove the example from the dataset of record so the next training run excludes it. Version datasets immutably, and record which model versions were trained on which dataset versions.

Eval and golden sets

Evaluation sets are often hand-curated from real conversations and then forgotten. Build them from masked or synthetic data. If a real conversation is used, it carries the same subject ID and is purged with everything else.

Model weights

There is no reliable, auditable way today to remove one individual's influence from trained weights. The defensible engineering position is prevention: do not fine-tune on identifiable personal data where a masked version will do, and retrain on a fixed cadence from a cleaned dataset. Add output filters that block known identifiers of erased subjects as a compensating control.

An end-to-end erasure workflow for AI products

  1. Intake. Publish the request channel and the identifiers you need — username, email, mobile, customer ID — as Rule 14(1) requires. Verify the requester against an identifier you already hold.
  2. Resolve identity graph. Map the person to every internal subject ID, tenant ID and pseudonymous token across systems.
  3. Fan out deletes. Emit an erasure.requested event consumed by each store owner: primary DB, object storage, vector DB, cache, log pipeline, dataset registry, analytics warehouse.
  4. Processors. Call each processor's deletion API or send a written instruction — LLM API providers, observability tools, cloud hosts. Section 8(7) makes this your duty, not theirs.
  5. Retention routing. Move records that Rule 8(3) requires you to keep into the restricted archive with a per-record expiry date.
  6. Verify and evidence. Each consumer returns a signed completion record; the orchestrator stores the evidence pack.
  7. Respond. Reply to the Data Principal with what was erased, what is retained and until when, and the Rule 9 contact details — inside your published grievance window, which cannot exceed 90 days.

Store hashes and counts in the evidence pack, not content.

Common mistakes in LLM erasure

  • Deleting the document but not the embeddings. The source file is gone, yet retrieval still surfaces its chunks.
  • Forgetting processor copies. Hosted LLM APIs, tracing tools and labelling vendors may retain prompts. Check retention settings and contract terms; Rule 6(1)(f) expects security clauses in processor contracts.
  • Hard-deleting logs immediately. Wiping all traces on request can conflict with Rule 8(3)'s one-year minimum and leave you unable to investigate a breach.
  • Holding everything forever "for compliance". The reverse error. Rule 8(3) says retain for a minimum of one year and then erase, unless another law requires more.
  • No subject tagging in training data. Without lineage you cannot prove an erased person is excluded from the next fine-tune.
  • Free-text identity in eval sets. Copies of real chats in spreadsheets escape every deletion job.
  • Treating embeddings as anonymous. If metadata or neighbouring chunks re-identify the person, the vector is personal data.

Large platforms in the Third Schedule — e-commerce and social media entities with at least 2 crore registered users, and online gaming intermediaries with at least 50 lakh — also face automatic erasure after three years of inactivity, with a 48-hour prior notice. See Rule 8 retention and erasure.

What to build before 13 May 2027

Roughly six months remain before Rules 8, 9 and 14 apply. A sensible sequence for an AI team:

  • December–January: inventory every AI data store and processor; add subject IDs to vector metadata, traces and datasets going forward.
  • February: backfill subject tags on existing indexes, or schedule a re-ingestion that adds them.
  • March: build the erasure orchestrator and the restricted retention archive with per-record expiry.
  • April: run end-to-end drills — erase a test user and prove zero results across every store; test processor deletion APIs.
  • May: publish the request channel, identifiers, Rule 9 contact and grievance timeline.

Assign one owner for the erasure orchestrator, so that any AI data store added later — a new index, cache or eval bucket — is wired into the delete path by default rather than discovered during a request.

Pair this with your consent withdrawal design, because withdrawal triggers the same erasure path under section 8(7). The UX side is covered in consent and withdrawal UX, and the broader rights framework in Data Principal rights under Rule 14.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Tag every vector chunk with subject and source IDs

    Store subject_ids, source_doc_id and tenant_id as filterable metadata so a person's vectors can be found and deleted.

    DPDP Act s. 12

  2. Schedule index compaction after deletes

    Ensure tombstoned vectors are physically removed on a fixed cadence rather than lingering in segments.

    DPDP Act s. 8(7)

  3. Maintain training-data lineage

    Map every fine-tuning and eval example to its source record so erased subjects are excluded from the next run.

    DPDP Act s. 8(7)

  4. Build a restricted retention archive

    Hold Rule 8(3) records encrypted, unreadable by product features, with per-record one-year expiry.

    Rule 8(3)

  5. Instruct every processor to erase

    Use deletion APIs or written instructions for LLM providers, tracing tools and hosts, and keep confirmations.

    DPDP Act s. 8(7); Rule 6(1)(f)

  6. Publish request channel and identifiers

    State on your website or app how to request erasure and which identifiers are needed.

    Rule 14(1)

  7. Include the contact person in every response

    Every reply to a rights request must carry DPO or contact-person business details.

    Rule 9

Key dates

  1. DPDP Rules notified; Board provisions in force

  2. Consent Manager registration (Rule 4) in force

  3. Rules 8, 9 and 14 and core erasure procedures apply

Answers

Frequently asked questions

Do I have to retrain my model when someone asks for erasure under DPDP?

The Act does not mention model weights specifically, and there is no reliable way to remove one person's influence from trained weights today. The practical position is to remove the person from the training dataset of record so future runs exclude them, retrain on a regular cadence, and avoid fine-tuning on identifiable data in the first place. Output filters for known identifiers are a sensible compensating control.

Are embeddings personal data under the DPDP Act?

Embeddings are personal data when they relate to an identifiable individual, which is usually the case because vector stores keep metadata and source chunks alongside them. Treat a vector as personal data if it can be linked back to a person through its metadata, its source text or a lookup table, and include it in erasure.

Can I keep prompt logs after a user deletes their account?

Yes, to the extent Rule 8(3) requires. From 13 May 2027 every Data Fiduciary must retain personal data, traffic data and processing logs for at least one year from processing for Seventh Schedule purposes, even if the user deletes the account, as the Rule's e-book illustration shows. After that, erase them unless another law requires longer retention.

How long do I have to respond to an erasure request?

The DPDP Rules require each Data Fiduciary to publish its grievance response period, which must not exceed 90 days under Rule 14(3). Aim much lower for erasure: an automated orchestrator can complete deletions in days. The response must also include the business contact of your DPO or designated contact person under Rule 9.

Does the hosted LLM provider have to delete my users' prompts?

The provider is usually your Data Processor, and section 8(7) makes you responsible for causing your processors to erase. Check whether the provider retains prompts, for how long, and whether it offers a deletion API or zero-retention setting. Put erasure and security obligations in the contract, as Rule 6(1)(f) contemplates.

What about erasure in vector databases that only soft-delete?

A soft delete hides a vector from queries but may leave it in storage segments until compaction. Treat erasure as complete only after compaction or segment rewrite has run, schedule it regularly, and apply the same timetable to index snapshots and replicas kept for rollback.

Legal references and sources

  • Digital Personal Data Protection Act, 2023 — Sections 8(7), 12 and 13 (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 8(3) and Seventh Schedule (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 14 (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 9 (official text)

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.