Build Your Own Research Assistant — Projects, Notebooks and RAG

In this session

• 01

  The opening mirror, and what it teaches: a realistic junior's prompt about a live matter goes up — names, amounts, a diagnosis-shaped fact — and the room votes on whether it is safe to send. Then the identifiers are highlighted, and the abstraction move is performed live: the same question asked about 'a licensee' and 'a sum' gets the same quality of answer with nobody's client in it. The mechanism behind the rule: consumer free tiers log conversations and, by default, most train on them — the opt-out switches exist and the lab flips them — and a workspace makes the stakes permanent, because what you upload persists by design.

• 02

  RAG, drawn once and drawn properly: your documents are stored and indexed; at each question, a retriever fetches the passages that look most relevant; those passages go into the window; the model answers from them, ideally with a pointer to where. Every strength follows (the corpus can be far bigger than any window; answers come with addresses; the model is reshaping, not recalling) — and every failure mode is a clerk's failure: the retrieval miss (the right passage never fetched, so the answer rests on the wrong ones), the wrong pointer (a citation to a passage that says something adjacent), and the override (the model answering from its training anyway, especially when the sources contradict what it 'remembers'). Grounded is better; grounded is not true: the Stanford benchmark of professional legal-research tools found even citator-backed systems hallucinating on roughly one query in six.

• 03

  The harness zoo, demystified — four artefacts, one mechanical question each: what persists? Custom instructions: a standing preamble to every chat (free on the major tiers; small; the right home for 'always cite paragraph numbers; flag anything unverified'). A Project (ChatGPT: free tier, five files per project; Claude: free tier, five projects): instructions plus files plus chats, in one container — a small RAG over your bundle. A Gem (Gemini, free): a saved instruction set with optional knowledge files. NotebookLM (free, 50 sources per notebook, each source up to half a million words): the purest of the four — source-grounded by design, every answer carrying inline citations that click through to the passage, and, unusually for a free consumer tool, not training on your uploads. One closed door named honestly: building a custom GPT is no longer a free-tier exercise (at the time of checking, the vendor's own pages disagreed about which paid plans still allow it) — the free-tier route is Projects, Gems and notebooks, and it is enough. One opened door worth knowing: ChatGPT Projects can now be shared, even on free accounts — a moot team can work one workspace together — with the caveat that a shared project trains on members' data unless every member has turned training off.

• 04

  The instruction set is where a law student's training shows. The workspace's standing orders, written like office policy: answer only from the sources in this workspace; cite the document and paragraph for every proposition; where the sources do not address the question, say 'not in the sources' — do not supplement from general knowledge; where sources conflict, say which is later and which is higher; maintain the running research log when asked. Each line exists because Session 1 taught you what the machine does without it. This is harness engineering: not code, but policy — the same skill as drafting instructions for a junior who is brilliant, literal, and constitutionally incapable of saying 'I don't know' unless ordered to.

• 05

  The live build, on the bundle the course has been assembling all along: State v. Navjot Sandhu, Anvar P.V. v. P.K. Basheer, Shafhi Mohammad, Arjun Panditrao Khotkar, and the text of s.63 of the Bharatiya Sakshya Adhiniyam — the electronic-evidence line, which is simultaneously a doctrine-tracing exercise (watch the certificate requirement fall, rise, waver and settle across four cases, with pinpoints) and the exact law Session 7 will need. Then the trap question — one the bundle cannot answer — and the workspace, under its instruction set, says 'not in the sources', which is the most professionally reassuring sentence a machine can produce.

• 06

  Then the moment the session is really for: the staleness catch. The workspace is asked who may sign the s.63(4) expert certificate — and answers from its newest source as if the law stopped there, because for this workspace, it did. The room is then shown Pune Bar Association v. Union of India (SC, 22 May 2026), which upheld the certificate regime and read 'expert' beyond the notified s.79A examiners; the order is added to the bundle live; the same question is re-asked; the answer changes, with a new pinpoint. A workspace is a snapshot, not a subscription: it knows nothing it was not given, including last month's law — updating the bundle is now part of updating the matter, and 'is my bundle current?' joins 'is it still good law?' in the standing checklist.

• 07

  What never goes in, now as doctrine rather than caution: a consumer workspace is a standing disclosure to a third party, so client documents, instructions and identifying facts stay out — full stop. The legal frame, previewed here and completed in Session 7: privilege under s.132 BSA belongs to the client and is not yours to waive into a training corpus; the professional rules make the client's confidence your standing duty; the DPDP Act — its substantive obligations commencing May 2027 — will make the advocate a data fiduciary engaging an unbriefed processor the moment client data is pasted; and In Re: Summoning Advocates (2025 INSC 1275) has already shown the Supreme Court thinking hard about exactly what a shared repository of a lawyer's files exposes. The working rule the course will not soften: published law in, clients never — and the enterprise-grade exceptions are a thing firms buy, not a thing free tiers give.

• 08

  When not to build one: a workspace earns its keep on recurring work over a stable bundle — a moot, a seminar paper, an internship's running matter, exam revision over a course's readings. For a one-off question, Session 3's paste-and-ask beats it. The diagnostic vocabulary grows one more word: alongside the prompt ceiling and the model ceiling there is a tool ceiling — and last session's two-hundred-page judgment, which defeated the free chat window, fits inside a single free NotebookLM source with room to spare. Sometimes the answer is not a better model; it is a better-shaped tool, and it is free.

The build-along

Build it with the room. Leave holding it.

The class constructs the thing alongside the presenter — the prompt, the workspace, the pipeline — with the mechanism explained as it is built, and a named skill at the end.