Knowledge base and retrieval

The knowledge base is where your organization's source-of-truth documents live so Henry can use them. It is the difference between an AI that drafts generic feedback and an AI that drafts feedback aligned to your actual leveling guide, your actual competency model, and your actual process.

This article covers what the knowledge base is and why it exists, how to upload and categorize documents, how indexing works, how retrieval shows up in Henry's answers, how to remove or replace documents, and best practices for keeping the knowledge base useful over time.

Plan availability: Henry Agent is available on the Agentic plan. Team plan customers can upgrade in Settings → Billing.

What the knowledge base is

The knowledge base is a collection of documents your organization has uploaded and made available to Henry. When Henry needs background to answer a question or draft content, it searches the knowledge base for relevant snippets and uses them as grounding for its response.

Concretely, the knowledge base solves three problems:

• Henry does not know your company's specific process unless you tell it. Generic AI knows what a 1:1 is in general; only your knowledge base knows what your company's expectations for a 1:1 are.
• Your leveling and competency frameworks are unique. "Senior engineer" means different things in different orgs. Without your leveling guide, Henry's drafts default to generic phrasing.
• Process and policy answers should match the source of truth. When someone asks "what is the policy on mid-cycle promotions," Henry should pull the answer from your handbook, not invent one.

The knowledge base is gated by two feature toggles: aiKnowledgeBase (organization-wide, controlled by your admin) and the underlying Henry availability. If your organization has Henry but not the knowledge base feature, Henry still works for drafting and summarization but does not have document context to draw on.

Who manages the knowledge base

Knowledge base administration is admin-only. Org admins upload, categorize, replace, and remove documents. Non-admin users do not see the knowledge base management UI, but they benefit from it implicitly — Henry's answers improve once useful documents are indexed, regardless of who triggered the question.

This separation is intentional. The knowledge base shapes everyone's Henry experience, so it should be curated deliberately. A free-for-all where every user uploads documents would quickly produce a noisy, conflicting corpus that hurts answer quality.

Uploading documents

Documents are uploaded from the admin knowledge base management view, reachable from Settings → AI → Knowledge base (admin role required).

Supported file types

File type	Notes
PDF	The most common format. Text is extracted and indexed.
DOCX	Microsoft Word documents. Text and basic formatting are extracted.
Images	PNG, JPG. Used for diagrams or scanned pages. Optical recognition extracts text where present.

Other file types — spreadsheets, slide decks, plain text, Markdown — are not currently supported through the upload UI. If you have content in another format, export it to PDF or DOCX before uploading.

Document types and categories

When you upload a document you assign it a type. The available types are:

• Leveling guide — career frameworks, level expectations, role definitions.
• Competency model — competencies, behaviors, rubrics for evaluation.
• Process doc — how-to documents, runbooks, procedure guides (e.g., how the promotion process runs).
• Other — anything else relevant to performance management context.

The type is used at retrieval time. When Henry is helping with a summary, it weighs leveling guides and competency models more heavily. When Henry is answering a process question, it leans on process docs. The type tells Henry what shape of question this document is most likely to answer.

You also give the document a name. Choose something descriptive — "Engineering Leveling Guide v3 (2026)" is a far more useful citation than "guide.pdf."

Visibility and scope

All documents in the knowledge base are organization-scoped. Anyone in your organization who interacts with Henry benefits from the indexed content. There is no per-team or per-document gating today.

This means you should not upload documents that contain sensitive information meant for a narrow audience. The knowledge base is for content you would be comfortable having Henry reference in any user's interaction inside your org.

Documents from one organization are never available to another organization, on any tier.

How indexing works

Once you upload a document, it does not immediately become available to Henry. It goes through an indexing pipeline.

The status lifecycle

A document moves through these states:

1. Pending — the upload completed and the document is queued for processing.
2. Processing — the document is being parsed, chunked, and embedded.
3. Indexed — the document is fully indexed and available to Henry's retrieval.
4. Failed — something went wrong during processing. The admin UI shows the failure reason and offers a retry.

You can see the status of every document in the knowledge base management view. Until a document is Indexed, Henry does not see its content.

What happens during processing

Three things happen behind the scenes:

• Parsing extracts the text content from the file. PDFs and DOCX files are parsed for their text; images are run through optical recognition.
• Chunking splits the document into smaller pieces. Henry's retrieval works at the chunk level, not the whole-document level — this is what allows it to pull a specific paragraph rather than a 50-page handbook.
• Embedding turns each chunk into a vector representation that captures its meaning. At question time, Henry compares the question's embedding to the chunk embeddings to find the most relevant pieces.

You do not need to think about chunk sizes or embedding details — the pipeline handles it. The thing to know is that the unit of retrieval is a chunk, which is roughly a paragraph or two.

Indexing latency

Indexing is not instant. Most documents finish in a short window after upload, but very large documents may take longer. Plan ahead — if you are about to roll out a new leveling guide and want Henry to use it on Monday, upload it on Friday.

The status indicator in the admin view tells you whether a document is still processing. Until it shows Indexed, Henry will answer without knowledge of the new content.

How retrieval shows up in answers

When you ask Henry a question or it is generating a draft, the retrieval pipeline runs in the background. If it finds relevant chunks in the knowledge base, those chunks are passed to Henry as context. The answer Henry produces is grounded in that context, and the source documents are cited.

Inline citations

Citations appear as small inline references inside Henry's response. The citation list at the bottom of the message names the source documents.

Click a citation to see the snippet Henry pulled from. Click the document name to open the source if you have access. The snippet is the chunk Henry actually retrieved, not the entire document, so reading the snippet tells you exactly what context Henry had.

Example

You: What is our policy on mid-cycle promotions?
Henry: Mid-cycle promotions are allowed in two cases: (1) the employee was operating at the higher level for at least one full cycle before the request, and (2) the employee's current cycle has not yet started review submissions. Both must be approved by the skip-level manager and HR before the change is recorded. [1]

[1] Engineering Leveling Guide v3 (2026) — page 14

The bracketed reference is a citation. The full document name and the section context appear in the citation footer.

When Henry does not cite

Not every answer has a citation. Two reasons:

• The question was about your in-product data, not the knowledge base. "What did I write about Sara last quarter?" pulls from observations, not from documents, so there is nothing to cite.
• The retrieval did not find a strong match. Henry will still answer, but it will draw on its general knowledge rather than your documents. If you expected a knowledge base hit and did not get one, that is a signal that either the document is missing, the relevant content is buried, or the question was phrased in a way the retrieval did not match well.

If a question is one you want grounded in your knowledge base and Henry is not citing, check that the document is uploaded and indexed, and try rephrasing the question.

Two-tier retrieval

Behind the scenes, the retrieval has a stricter threshold for what is shown to you and a more permissive one for what informs the answer. This means Henry may use a chunk as soft context without surfacing it as a citation. The visible citations are the ones the system is confident enough about to display.

Removing or replacing a document

Documents can be removed from the knowledge base from the admin view. Removal is immediate — the document and its chunks are deleted, and Henry will no longer retrieve from it.

For replacement (e.g., publishing a new version of your leveling guide):

1. Upload the new version with a clear name that includes the version or year.
2. Wait for it to reach Indexed status.
3. Remove the old version.

This sequence keeps Henry grounded continuously. If you remove the old version before the new one finishes indexing, there will be a window where Henry has no leveling guide to draw on.

If you want to preserve old versions for reference outside Henry, archive them in your document storage system rather than keeping them in the knowledge base. Stale versions in the knowledge base produce inconsistent retrieval — Henry may pull from the old guide when it should be pulling from the new one.

Best practices

A few principles for keeping your knowledge base useful.

Keep documents current

Out-of-date documents produce wrong answers. When your leveling guide changes, replace it. When your promotion process changes, update the process doc. A knowledge base of last year's documents is worse than no knowledge base — Henry confidently cites obsolete content.

Set a cadence for review. A quick audit each cycle (or each quarter) catches documents that need refreshing.

Scope appropriately

The knowledge base is for documents that should inform Henry's answers in any context. If a document is highly specific to a narrow audience, leave it out. Examples of what belongs:

• Leveling guides, role definitions, competency models.
• Manager handbooks and IC handbooks.
• Process docs for promotions, performance improvement, calibration, 1:1s.
• Glossaries and terminology guides.
• Onboarding playbooks.

Examples of what to keep out:

• Sensitive personnel files.
• Drafts of policies that have not been published.
• Historical content that is no longer accurate.
• Long-form internal essays that are interesting but not authoritative.

The smaller the knowledge base, the easier it is to maintain quality. Resist the urge to upload everything.

Prefer source-of-truth over duplicates

If the same content exists in two documents (e.g., a leveling guide and a one-pager that summarizes it), pick the one you want to be authoritative and remove the other. Two versions of the same content compete during retrieval and produce inconsistent results.

Use clear filenames and types

Henry surfaces document names in citations. "guide.pdf" is unhelpful; "Engineering Leveling Guide v3 (2026)" tells the reader exactly what they are looking at. Same for types — putting a process doc into "Other" makes retrieval less precise. Tag it as Process doc.

Watch the citations

The fastest way to know whether your knowledge base is working is to read the citations on Henry's answers. If the cited snippets are relevant and accurate, the base is healthy. If Henry is pulling odd or outdated chunks, that is a signal to clean up.

You can also experiment by asking questions you know the answer to and checking whether Henry's citations match what you expect.

Limits

The knowledge base has limits on the number of documents and the total size of indexed content per organization. You can see your current usage in the admin view. If you approach the limit, plan for cleanup before adding new documents — removing stale versions is usually cheaper than expanding capacity.

Per-document size limits also apply. If an upload is rejected for size, split the document into smaller logical sections (e.g., one PDF per chapter) and upload each separately. This also tends to improve retrieval, since smaller documents make it more likely the right chunks surface.

Next steps

• Meet Henry explains the broader picture of what Henry knows about your organization.
• Using the Henry panel covers how citations appear in the chat UI.
• Henry analytics shows admins how often retrieval is being used and where it is most valuable.