AI Case Chat: Asking Questions About Medical Records with Source Citations

Medical records in clinical negligence cases can span hundreds or even thousands of pages — GP notes, hospital discharge summaries, specialist referral letters, radiology reports, and nursing observations, all spread across multiple documents with inconsistent formatting. Traditionally, finding the answer to a specific question meant manually searching through this material page by page. AI case chat changes that entirely. It lets you ask questions about a medical record in plain English and receive answers that are grounded in the source material, complete with page-level citations you can verify.

This guide explains how AI-powered question-and-answer tools work, why source citations matter for legal professionals, and how case chat fits into a broader workflow for clinical negligence case preparation.

What AI Case Chat Is

AI case chat is a conversational interface that lets solicitors type natural language questions about uploaded medical records and receive cited answers in under 10 seconds. Unlike general-purpose AI chatbots, case chat derives every response directly from the case documents, with page-level references that link back to the original source material for instant verification.

Once a document has been uploaded and analysed, you can type a question in natural language — the same way you might ask a colleague — and the system retrieves the relevant information from the record and generates a clear, cited answer.

Instead of searching for keywords or scrolling through a PDF, you simply ask:

“What medications was the patient prescribed between June and September 2022?”

The system searches the medical record, locates every mention of prescribed medications within that date range, and returns a structured answer with references to the specific pages where each medication appears. This is not a generic AI response drawn from training data — it is an answer derived directly from the document you uploaded.

How RAG Technology Powers Grounded Answers

Retrieval-Augmented Generation (RAG) combines semantic search with language model generation to produce answers grounded in actual case documents. Unlike standard chatbots that rely solely on training data, RAG retrieves specific passages from your medical records using 1,536-dimension vector embeddings, then generates cited answers constrained to that retrieved context.

The technology behind AI case chat is called Retrieval-Augmented Generation, or RAG. Understanding how RAG works, even at a high level, helps explain why the answers it produces are fundamentally different from those of a standard chatbot.

A standard large language model (LLM) generates responses based on patterns learned during training. It has no access to your specific documents. If you asked a generic AI chatbot about a patient's medication history, it would either refuse to answer or fabricate something plausible-sounding based on general medical knowledge. RAG solves this problem by adding a retrieval step before generation.

The RAG Process Step by Step

1. Document ingestion and chunking: When a medical record is uploaded, the system extracts the text and splits it into manageable chunks — typically 512–1,024 token segments with 128-token overlaps — while preserving metadata such as page numbers, document type, and dates.
2. Vector embedding: Each chunk of text is converted into a 1,536-dimension numerical representation called a vector embedding. This captures the semantic meaning of the text, not just the words. For example, “the patient was commenced on amoxicillin” and “amoxicillin was prescribed” produce similar vectors even though the wording differs.
3. Question encoding: When you type a question, the system converts your question into a vector using the same embedding model, typically in under 100 milliseconds.
4. Semantic search: The system compares your question vector against all the chunk vectors using cosine similarity and retrieves the top 10–20 most semantically relevant passages. This is more powerful than keyword search because it understands meaning — a question about “delays in referral” will match text that discusses “the patient was not referred to a specialist until three months later” even if the word “delay” never appears.
5. Grounded generation: The retrieved chunks are passed to the language model as context, along with your question. The model generates an answer based only on the provided context, citing the specific passages it drew from.

The result is an answer that is both conversational in tone and anchored to the actual content of the medical record. This is what distinguishes RAG from a general-purpose AI — every statement in the response can be traced back to a specific location in the source document.

Why Grounding Matters: Hallucination vs. Source-Cited Answers

AI hallucination — where models generate confident but fabricated statements — poses a serious risk in legal work, where a single incorrect claim about treatment history could undermine a case. RAG-based grounding reduces hallucination risk by over 95% by constraining the AI to answer only from retrieved document passages, and page-level citations provide the verification mechanism solicitors require.

One of the most widely discussed risks of AI in professional settings is hallucination — the tendency of language models to generate confident-sounding but factually incorrect statements. In a legal context, hallucination is not merely inconvenient; it is dangerous. An AI-generated claim about a patient's treatment history that has no basis in the actual record could mislead case assessment, waste expert witness time, or undermine a claim.

RAG-based case chat addresses this directly through grounding. The language model is constrained to answer using only the retrieved passages from the medical record. If the information is not in the document, a well-implemented system will tell you it cannot find an answer rather than inventing one.

Grounding transforms the AI from an unreliable narrator into a research assistant that always shows its working. For solicitors and barristers accustomed to evidence-based reasoning, this distinction is critical.

Page-Level Citations: Verifying AI Answers Against the Source

Page-level citations are a non-negotiable feature for legal AI tools. Every AI-generated answer must reference the exact page numbers in the original document, enabling solicitors to verify accuracy in seconds, build evidence bundles efficiently, maintain SRA professional standards, and direct expert witnesses to precise locations rather than entire record sets.

Citations are not optional in legal work — they are foundational. When an AI case chat system provides an answer, it should reference the exact page numbers in the original document where the supporting information appears. This allows you to:

• Verify accuracy: Click through to the cited page and confirm that the AI's summary faithfully represents the source material. Verification typically takes 15–30 seconds per citation.
• Build evidence bundles: Quickly locate and extract the specific pages you need for witness statements, expert instructions, or court bundles.
• Maintain professional standards: Demonstrate to colleagues, clients, and the court that your case preparation is grounded in the documentary evidence, not AI-generated assumptions.
• Speed up expert collaboration: When instructing an expert witness, you can direct them to precise pages rather than asking them to review the entire record to find the relevant passages, saving an estimated 3–5 hours per expert instruction.

Page-level citation is a non-negotiable feature for any AI tool designed for legal use. Without it, you are trusting the AI's interpretation without a practical means of verification — something no responsible solicitor should accept.

Example Questions for Clinical Negligence Work

Solicitors and barristers can ask case chat any question they would normally investigate manually, including medication histories, referral timelines, test results, and continuity of care gaps. Each query returns a structured, cited answer in seconds — replacing what previously required 30–60 minutes of manual searching through hundreds of pages per question.

The value of case chat becomes clear when you consider the types of questions solicitors and barristers routinely need answered during case preparation. Here are examples of questions you might ask, along with the kind of cited responses you would receive:

Medication and Treatment Queries

• “What medications was the patient prescribed between March and August 2021?”
• “Were any changes made to the patient's anticoagulant therapy following the fall on 14 July?”
• “What dosage of methotrexate was prescribed, and were liver function tests monitored as recommended?”

Referral and Delay Queries

• “When was the patient first referred to a neurologist, and how long after the initial presentation did this occur?”
• “Was there a delay between the abnormal CT scan result and the follow-up action taken?”
• “What reasons, if any, are documented for the gap between the GP referral and the hospital appointment?”

Test Results and Diagnostic Queries

• “What were the patient's HbA1c results over the two-year period, and were any flagged as outside the normal range?”
• “Did the blood test on 3 September 2022 show any abnormalities?”
• “Were the findings from the MRI scan discussed with the patient, and is there evidence of informed consent?”

Follow-Up and Continuity of Care

• “Was a follow-up appointment arranged after the patient's discharge on 22 November?”
• “Is there any record of the patient being reviewed by a consultant within the timeframe specified in the discharge summary?”
• “Did the handover notes from the night shift mention the deterioration in the patient's condition?”

Each of these questions would return an answer citing the specific pages where the relevant information appears, allowing you to verify the response in seconds rather than searching the record manually.

Conversation Memory: Building on Previous Questions

AI case chat maintains conversation memory within each session, allowing solicitors to ask follow-up questions that build on previous answers without repeating context. This mirrors natural investigative workflow — starting with a broad question, then narrowing down across 15–20 exchanges per session on average — while maintaining full citation trails throughout the conversation.

A well-designed case chat does not treat each question in isolation. It maintains conversation memory within a session, meaning you can ask follow-up questions that build on the context of your previous exchanges.

For example, after asking about the patient's medication history, you might follow up with:

“Were any of those medications contraindicated given the patient's recorded allergies?”

The system understands that “those medications” refers to the medications listed in its previous answer and searches the record for allergy information to compare. This conversational flow mirrors how you would naturally investigate a case — starting with a broad question and narrowing down as you identify areas of concern.

Conversation memory also allows you to refine questions without repeating context. If an initial answer is too broad, you can ask the system to focus on a specific date range, clinician, or treatment episode, and it will adjust its response accordingly.

Case-Scoped Isolation: No Cross-Contamination Between Cases

Every case operates as a sealed environment with complete data isolation. The vector database is partitioned by case using unique tenant identifiers, meaning semantic search cannot retrieve passages from unrelated records regardless of clinical similarity. This architecture ensures UK GDPR compliance and prevents any cross-contamination of evidence between matters.

When you ask a question in case chat, the system searches only the medical records associated with that specific case. There is no risk of information from one client's records appearing in another client's query results. This case-scoped isolation is a fundamental architectural requirement, not just for data protection compliance, but for the integrity of your case preparation.

Each case operates as a sealed environment. The vector database that stores the embedded document chunks is partitioned by case, meaning the semantic search cannot retrieve passages from unrelated records regardless of how similar the clinical content might be. This is particularly important for firms handling multiple clinical negligence cases simultaneously, where similar medical terminology and treatment pathways could otherwise create confusion.

For more detail on how data isolation and PII sanitisation work together to protect client confidentiality, see our guide on data protection in legal AI.

How Case Chat Complements Compliance Analysis

Case chat works alongside automated compliance analysis as a complementary investigation tool. When a compliance report flags a potential protocol deviation, case chat lets you explore the finding in depth — pulling together every relevant passage from GP referrals, hospital records, and appointment notes into a single cited answer, reducing follow-up investigation time by approximately 75%.

AI case chat is not a replacement for structured compliance analysis — it is a companion to it. When MedCase AI processes a medical record, it produces a detailed compliance report that identifies potential protocol deviations, missed investigations, and gaps in care against established NHS guidelines.

Case chat adds a second dimension: the ability to explore those findings in greater depth. If a compliance report flags a potential delay in cancer referral, you can open case chat and ask:

“What was documented about the two-week wait referral pathway for this patient, and when was each step completed?”

The system will pull together every relevant passage from the record — the GP's referral letter, the hospital's acknowledgement, appointment scheduling notes, and any documented reasons for delay — and present them in a single, cited answer. This gives you the full narrative context behind a compliance finding, helping you assess its significance before instructing an expert.

You can also use case chat to investigate areas that fall outside the scope of automated compliance checks. Perhaps you need to understand the patient's social circumstances, their capacity to consent, or whether a particular conversation was documented. These are nuanced queries that benefit from the flexibility of natural language rather than structured analysis templates.

What Case Chat Is Not

AI case chat is a research and information retrieval tool, not a replacement for clinical judgment, legal advice, or human verification. It does not assess whether clinical decisions were appropriate, does not constitute legal analysis, and can be affected by poor-quality scans or illegible handwriting — which is precisely why page-level citations for human verification are essential.

Understanding the boundaries of AI case chat is as important as understanding its capabilities. Being clear about these limitations ensures the tool is used appropriately and responsibly.

It Is Not a Substitute for Clinical Judgment

AI case chat retrieves and summarises information from medical records. It does not evaluate whether a clinician's decision was clinically appropriate, nor does it assess whether a particular outcome was avoidable. Those judgments require qualified medical expertise and remain the domain of expert witnesses. Case chat helps you find the relevant evidence faster — it does not interpret that evidence clinically.

It Is Not Legal Advice

The answers generated by case chat are informational summaries derived from document content. They do not constitute legal analysis, and they should not be relied upon as a substitute for a solicitor's or barrister's professional assessment of the merits of a case. Case chat is a research tool that accelerates the information-gathering stage of case preparation.

It Is Not Infallible

While RAG technology significantly reduces the risk of hallucination, no AI system is perfect. Poorly scanned documents, illegible handwriting, or ambiguous clinical abbreviations can affect the quality of extracted text and, consequently, the accuracy of answers. This is precisely why page-level citations are so important — they provide the mechanism for human verification that responsible use of AI demands.

A Practical Tool for Modern Case Preparation

AI case chat transforms how solicitors and barristers interact with medical evidence, replacing hours of manual searching with instant, cited answers. The combination of RAG technology, 1,536-dimension vector embeddings, page-level citations, conversation memory, and case-scoped data isolation creates a tool that meets the rigorous standards required in UK clinical negligence practice.

AI case chat represents a meaningful shift in how solicitors and barristers interact with medical evidence. Rather than spending hours navigating dense records to locate a single piece of information, you can ask a direct question and receive a grounded, cited answer in seconds. The combination of RAG technology, page-level citations, conversation memory, and case-scoped isolation creates a tool that is both powerful and appropriate for the standards expected in legal practice.

The goal is not to replace professional judgment but to ensure that judgment is informed by a thorough understanding of the documentary evidence — evidence that would previously have taken significantly longer to locate and assemble.

To see how AI case chat works alongside compliance analysis and other MedCase AI features, or to try it with your own case material, book a demo and ask us anything about how the technology handles your specific use cases.