How to build a legal knowledge base with Claude without uploading the entire firm

“Connecting the firm’s knowledge” sounds like a bulk-upload project. In reality, it’s a project of selection, governance, and retrieval. Technology can find passages and draft answers, but the firm must decide which materials are authoritative and for what task.

This framework can be used to build a knowledge base in Claude Projects or to prepare for a later integration. It does not, by itself, authorize the processing of confidential information and does not replace legal, contractual, privacy, or security review.

# What is a legal knowledge base for Claude?

It is a governed set of documents and instructions that Claude can use as context for specific tasks. It may include manuals, internal criteria, templates, approved clauses, FAQs, authorized internal precedents, and selected public sources.

It is not the same as a document management system. The document management system retains the case file and its history; the AI knowledge base presents a prepared selection for search and transformation. It also does not replace the case management system, version control, or specialized legal research.

Anthropic describes Projects as workspaces with history, instructions, and knowledge. When the volume approaches context limits, the product may use retrieval-augmented generation to find relevant content. That technical capability does not determine whether a source is current, applicable, or permitted.

# Why shouldn’t you upload the entire collection?

A bulk upload combines documents with different purposes, permissions, and validity periods. The result may look complete, but it becomes harder to know which version supported an answer and who was authorized to view it.

The most common risks are:

• mixing matters or clients that need to be kept separate;
• including drafts, emails, or notes without authority;
• retaining contradictory versions;
• exposing unnecessary personal data or secrets;
• retrieving a clause from the wrong jurisdiction or area of law;
• making it difficult to remove a document;
• increasing review costs without improving the answer;
• creating the false impression that “the whole firm” is available.

A broad knowledge base is not necessarily a useful one. For AI, retrieval quality depends as much on selection and structure as on volume.

# What scope should you start with?

Start with an operational question. For example: “Can the team prepare an initial comparison of confidentiality clauses against an approved guide?” This scope makes it possible to identify the materials, users, and expected outcome.

An initial module may contain:

1. a guide to approved criteria;
2. an output template;
3. reviewed examples;
4. a brief glossary;
5. essential public or internal sources;
6. rules for flagging missing information and escalating issues.

Avoid creating a knowledge base “to answer any legal question.” This definition makes it impossible to verify coverage and increases the risk that the tool will respond outside its authorized domain.

# How should sources be inventoried and classified?

Before uploading files, create an inventory outside Claude. Keep the inventory as a control record that does not depend on the provider’s interface.

Claude’s knowledge base should not become the only place where a document exists. The source of truth remains in the firm’s approved system; the knowledge base is a controlled representation for a specific use case.

# What documents should be excluded at the outset?

Exclude any material whose usefulness has not been demonstrated or whose processing is not authorized. In particular, avoid complete case files, client secrets, identification documents, sensitive data, documents subject to ethical walls, unreviewed drafts, and files with no owner.

It is also a good idea to exclude redundant sources. If five presentations repeat a policy with minor differences, the knowledge base needs one authorized version, not five possibilities. Diversity is useful for examples; ambiguity about which source is authoritative is not.

The LFPDPPP requires, among other principles, purpose limitation, proportionality, and accountability. Applied to design, this favors loading only the information needed for the documented purpose and establishing controls appropriate to its sensitivity.

# How should documents be prepared before uploading them?

A file that is readable by a person is not always reliably retrievable. Check text quality, scanning, headings, tables, appendices, and internal references. A PDF containing only images may require OCR and validation.

Preparation should:

• remove irrelevant pages and unnecessary data;
• preserve titles, sections, and numbering;
• separate documents with different purposes;
• include the version, date, and status in the visible content;
• standardize names without losing the source’s identity;
• check that the extracted text matches the original;
• record any transformation.

Don’t obscure provenance behind names like “final document 2”. Traceability starts before upload.

# How should Projects and permissions be organized?

Separation should follow the boundaries of the work. A Project may correspond to a practice area, a process, or an authorized set, but it should not group clients just because they use the same type of contract.

Anthropic publishes options for sharing Projects with specific people or with the organization, as well as permissions for use and editing. The firm should grant the least access necessary and reserve the editing of instructions and knowledge for designated owners.

A simple design distinguishes:

• administrator: manages members and settings;
• legal content curator: approves sources and instructions;
• user: consults and runs the workflow;
• reviewer: evaluates results and logs errors;
• process owner: decides on changes and continuity.

Access to the Project should not be granted for convenience. It should be based on the role and revoked when responsibilities change.

# What instructions does the knowledge base need?

Instructions turn the collection into a system for answering questions. They should specify which sources take priority, what output is expected, and what to do when there is insufficient support.

Include rules such as:

• answer only within the defined scope;
• cite the document and section used;
• distinguish source text from inference and recommendation;
• identify conflicts between versions;
• do not fill in missing facts or authorities;
• request information when a jurisdiction or date is needed;
• escalate when the question exceeds the module’s scope;
• remind users that a lawyer must review the output.

Instructions cannot fix a disorganized knowledge base. If two authorized sources contradict each other, the issue must be resolved in the inventory or made explicit through a priority rule.

# How do you test retrieval before using it?

Build a set of questions with known answers. It should include direct queries, alternative wording, questions with missing information, conflicts, and out-of-scope cases.

Evaluate at least:

Records both correct answers and false positives. A well-written answer that uses the wrong source must be classified as a failure.

# How do you keep the knowledge base up to date?

Every source needs a life cycle. When a new version is approved, the curator must remove or mark the previous one, update the inventory, and repeat the affected questions.

Schedule reviews based on risk and how quickly things change. An operational template can be reviewed when the process changes; a regulatory reference may need more frequent monitoring. Do not change the document date without a substantive change.

The maintenance log should state:

• what changed;
• who approved it;
• which source it replaces;
• which test cases were repeated;
• which users need to be informed;
• whether the change affects the permitted scope.

# When should you move from Projects to an integration?

Projects can work when people ask questions and review results in Claude. An integration may be necessary when there are many sources, dynamic permissions, retrieval from a content management system, structured outputs, or automatic traceability.

An integration does not fix weak governance. Before building one, the firm should have demonstrated what content it needs, how it is evaluated, and which permissions must be respected. Automating a knowledge base without an inventory only makes the disorder spread faster.

To test the workflow before expanding the knowledge base, see how to run a Claude pilot at a law firm.

# Frequently asked questions about legal knowledge bases in Claude

# Does Claude memorize all the documents in a Project?

Documents serve as Project knowledge and can be retrieved to provide context. That does not mean every answer uses all the material or that retrieval is infallible; it must be tested with questions and references.

# Can you use one Project per client?

That may be an option if the contract, policy, and permissions authorize it. Logical separation does not replace a review of confidentiality, retention, access, and content removal.

# How many documents should you upload first?

The minimum needed to answer the use case and run a representative sample. There is no universal number; the standard is verifiable coverage, not volume.

# What happens if a regulation changes?

The owner must update the source, remove or mark the previous version, and repeat the affected tests. The answer should not depend on the user remembering to warn Claude.

# Does a knowledge base prevent hallucinations?

No. It can reduce answers without context, but omissions, incorrect inferences, or imprecise citations can still occur. Human review, testing, and limits remain in place.

# Is giving files clear names enough?

No. You also need status, version, owner, classification, scope, permissions, and a record outside the platform.

# Conclusion: a useful knowledge base is smaller, clearer, and easier to govern

The goal isn’t for Claude to “know the whole firm.” It’s for it to assist with a task using authorized, current, retrievable sources, without crossing information boundaries.

Start with one module, keep a separate inventory, test difficult questions, and expand only when the evidence supports it. That discipline makes it possible to move from a collection of files to a controlled legal capability.

# Sources consulted

• Anthropic: What Projects are
• Anthropic: Create and manage Projects
• Anthropic: Claude Team
• Anthropic: Data use in commercial offerings
• Anthropic: Data retention for an organization
• Chamber of Deputies: LFPDPPP
• American Bar Association: Formal Opinion 512