What is Amon Hen?

Amon Hen is a project-scoped advisory assistant grounded in your team's own recorded knowledge. Unlike a generic AI chatbot, every response is anchored to decisions, rules, notes, and code context that you have explicitly captured for a specific project.

The name comes from Tolkien's Hill of Sight: a place of clear seeing. Amon Hen is designed to give you clear, authoritative answers drawn only from what your team has actually decided and recorded, not from inference or general knowledge alone.

What it is good for

• Answering questions like "What architectural decisions have we made?" or "What rules apply before I make a change?"
• Onboarding new team members with instant access to recorded team knowledge
• Maintaining decision history and rationale so institutional memory is never lost
• Getting consistent, context-grounded answers rather than generic AI guesses

What it is not

• A general-purpose chatbot: responses are scoped to your project context
• A document search tool: knowledge must be explicitly added to be used
• A replacement for your own judgment: it advises, it does not decide

LLM & RAG Comparison

How Amon Hen differs from a standard LLM

A standard large language model (LLM) such as ChatGPT or Copilot generates responses from patterns learned across billions of documents on the internet. It has broad general knowledge but knows nothing about your team, your decisions, your codebase, or your business rules unless you paste that context into every conversation manually. Each session starts fresh with no memory of what was said before.

Amon Hen wraps an LLM but fundamentally changes what it reasons from. The model never operates from general web knowledge alone. Every advisory response is generated exclusively from the structured, typed knowledge your team has explicitly recorded for a specific project. Your rules stay rules. Your decisions stay decisions. Nothing is hallucinated in to fill gaps.

How Amon Hen differs from RAG

Retrieval-Augmented Generation (RAG) is a technique that pairs an LLM with a search index. When a question arrives, RAG retrieves chunks of text from a document corpus that appear semantically similar to the query, then asks the LLM to answer using those chunks as context. It is commonly used to make LLMs answer questions about large document libraries: wikis, PDFs, internal knowledge bases.

Amon Hen takes a different approach for a different problem. RAG is good at surfacing relevant passages from large unstructured document sets. Amon Hen is built for structured, authoritative team knowledge where precision and type matter more than recall volume.

The practical result: if your team has decided that all database queries must go through the data access layer, Amon Hen will state that as a rule every time it is relevant, with the rationale your team recorded for it. A general LLM will offer patterns from common practice. A RAG system will retrieve whatever passage in your documents most resembles the query, which may or may not be the actual policy.

Core Concepts

Getting Started

Signing in

Amon Hen requires authentication. Click Sign In on the landing page. You will be directed to a secure login screen. After authenticating, you are returned to the advisor automatically.

Requesting access

If you do not yet have an account, click Request access on the landing page. Fill in your name, email address, and select whether you need Full user access (can ask questions and add knowledge) or Viewer only access (can read and ask, but not modify context). An optional note lets you explain your use case. The administrator will provision your account and send you a setup link.

First steps after signing in

1. Create or select a project in the left sidebar.
2. Browse Starter Packs to seed the project with pre-written knowledge (optional but recommended).
3. Add your own knowledge items: rules your team follows, decisions that have been made, and important context.
4. Ask questions using the Ask tab.

Projects

A project is the fundamental unit of organization in Amon Hen. All knowledge items and advisory responses are scoped to a single project. You can have multiple projects: one per system, team, or initiative.

Selecting a project

Use the Project dropdown in the left sidebar. Choosing a project loads its knowledge summary and makes it the active context for questions and new items. The sidebar shows a count of each knowledge item type the project contains.

Creating a project

1. Click + New Project below the project dropdown.
2. Type a descriptive name (up to 80 characters).
3. Press Enter or click Create.

The new project is immediately selected and ready for knowledge items.

The sidebar knowledge summary

When a project is selected, the sidebar shows two panels:

• Knowledge: a count of each item type (Rules, Decisions, Notes, Code Context). This tells you at a glance how populated your project is.
• Packs: any Starter Packs that have been applied to this project.

Deleting a project

Project owners see a small × button next to the project name in the sidebar. Click it to begin deletion. A confirmation prompt appears: you must confirm before anything is removed.

Asking Questions

The Ask tab is the primary interface for getting advisory responses. Select a project, type your question, and the advisor responds based entirely on the knowledge items recorded for that project.

How to ask a question

1. Make sure a project is selected in the left sidebar.
2. Click the Ask tab (active by default).
3. Type your question in the text box at the bottom of the screen.
4. Press Enter or click Send.

Press Shift+Enter to add a line break within a question without sending.

Suggestion chips

When you select a project, a row of suggestion chips appears. These are contextually generated starting points:

• Key decisions?
• Rules I should follow?
• What's established on architecture?
• What to know before making a change?

Click any chip to populate and send that question instantly.

Conversation context

Amon Hen maintains the last three exchanges (six messages) as conversation history. This means you can ask follow-up questions that reference your previous exchange:

Conversation history resets when you switch to a different project.

Response actions

Each advisory response includes three action buttons beneath it:

• Copy: copies the full response text to your clipboard.
• Thumbs up: marks the response as accepted. Used to score the advisory profile.
• Thumbs down: marks the response as rejected. Used to reduce the profile's score.

Writing effective questions

• Be specific: "What rules govern how we handle null values in ETL pipelines?" is better than "Rules?"
• Use the domain language your team uses: the advisor understands the terminology you have recorded.
• Ask for rationale when you want the why: "Why did we choose Databricks over Snowflake for transformation?"
• Ask multi-part questions in one message: "What are our GAAP close rules and who is responsible for the reconciliation step?"

Adding Knowledge

The Add Knowledge tab lets you populate a project's corpus. Every item you add is stored with a type, content, optional rationale, and a timestamp. The quality of advisory responses is directly proportional to the quality and completeness of the knowledge you add.

How to add an item

1. Select a project.
2. Click the Add Knowledge tab.
3. Choose a Type from the dropdown.
4. Enter the Content: the statement itself.
5. Enter a Rationale if applicable (required for decisions).
6. Click Add to Corpus.

Knowledge item types

See the full Knowledge Item Types reference for detailed guidance and examples for each type.

Giving Feedback

Your feedback directly improves the advisory engine. After every response, you can rate it with a thumbs up or thumbs down. Over time, these ratings determine which advisory profile is used to answer your questions.

How feedback works

• Thumbs up: the response was accurate, useful, and grounded in the project context as you understand it. Score contribution: 100 points.
• Thumbs down: the response was inaccurate, missing critical context, or otherwise not useful. Score contribution: 0 points.

Feedback is recorded against the advisory profile that generated the response. Once a profile accumulates 10 samples, it competes with other profiles for active status. The profile with the highest average score above that threshold is promoted and used for all subsequent questions.

Viewing feedback outcomes

Administrators can see aggregate feedback data on the Performance Dashboard, including acceptance rates, average scores per profile, and 30-day outcome distributions.

What Are Starter Packs?

A Starter Pack is a curated, versioned bundle of knowledge items pre-written for a specific technology, framework, discipline, or domain. Instead of manually entering dozens of rules and decisions from scratch, you can apply a Starter Pack to seed a project with authoritative baseline knowledge in seconds.

Starter Packs are maintained and versioned by the Amon Hen administrator. When a newer version of a pack is published, the old version remains applied to any projects that used it: you opt in to updates explicitly.

What a pack contains

Each pack contains a mix of knowledge item types appropriate to the domain:

• Rules: standard practices and constraints for the technology
• Decisions: common architectural choices with recorded rationale
• Notes: contextual background and reference information
• Code Context: naming conventions, patterns, and implementation guides

Pack verticals and technologies

Packs are organized along two dimensions, accessible from the sidebar on the Starter Packs page:

• Vertical: the discipline the pack addresses. Current verticals include Technical (engineering and data platform knowledge) and Finance (GAAP, IFRS, SEC reporting, audit standards, etc.).
• Technology: the specific platform or tool (BDC, Snowflake, Databricks, SAP, and others).

Selecting a vertical and technology in the sidebar filters the pack grid to show only relevant packs. The counts update dynamically as you filter.

Browsing & Filtering Packs

Navigate to Starter Packs from the header link on the main page. You must be signed in to access this page.

The sidebar filters

The left sidebar has two filter sections separated by a divider:

1. Vertical: All, Technical, Finance (and any future disciplines). Click a vertical to restrict the grid to packs in that discipline.
2. Technology: All, plus each technology group present in the current vertical filter. Click a technology to further narrow the grid.

Filters are AND-combined: selecting Finance + Snowflake shows only packs that are in the Finance vertical and the Snowflake technology group.

The count next to each filter item shows how many packs match that filter within the context of your other active filter. For example, the Vertical count reflects how many packs match within the currently active Technology filter.

Search

The search box below the sidebar filters the pack grid in real time against pack names and descriptions. Search is case-insensitive and works alongside the sidebar filters.

Pack cards

Each pack is displayed as a card. Cards are collapsed by default: only the pack name and version are visible. Hover over a card to expand it and reveal the description and action buttons.

A green dot in the top-left corner of a card indicates this pack has already been applied to the currently selected project. Applied packs are sorted to the top of the grid for quick reference. The date applied is shown on expanded applied cards.

Applying a Pack

Step-by-step

1. On the Starter Packs page, select the project you want to seed from the Apply to project dropdown at the top.
2. Browse or search for the pack you want.
3. Hover over the pack card to reveal the Apply button.
4. Click Apply. A loading indicator appears while the pack items are inserted.
5. A toast notification confirms success and shows the number of items added.

What happens when you apply a pack

All items in the pack are copied into your project's knowledge corpus exactly as written. They become your project's knowledge: you can add to them, and the advisor uses them immediately for all subsequent questions.

Applying the same pack twice to the same project is blocked: you'll receive a notification that it is already applied. To get items from a newer version of a pack, an administrator can upload the updated version, which will appear as a separate entry.

After applying

Return to the main Advisor and select the project. The sidebar will list the pack name under Packs. Start asking questions immediately. The pack's knowledge is already active in the context.

Access Requests Admin

The Admin page is accessible only to the Amon Hen administrator. The first panel, Access Requests, lists all users who have requested an account.

Request table columns

• Name: the name the requester entered
• Email: the email they will use to sign in
• Type: Full user or Viewer only
• Note: optional message from the requester
• Requested: date and time of the request
• Status: Pending, Approved, or Denied
• Actions: Approve or Deny buttons for pending requests

Filter by status

Use the filter buttons above the table (All, Pending, Approved, Denied) to focus on the relevant subset. Click ↻ Refresh to reload the table after taking actions.

Approving a request

Click Approve next to a pending request. The system creates the user account in Auth0 and generates a password-setup link. A modal appears with the link; copy it and send it directly to the user. The link is valid for 7 days.

Denying a request

Click Deny to mark the request as denied. No account is created. The requester is not automatically notified: notify them separately if appropriate.

Uploading Starter Packs Admin

The second panel on the Admin page lets you publish new Starter Packs or update existing ones without a code redeploy. Packs are defined in JSON and pasted directly into the upload form.

Using the format reference

The Required format block above the upload textarea shows the complete schema. Click Copy to copy it to your clipboard as a starting template.

Pack JSON structure

See the full Pack JSON Format reference below for field definitions and examples.

Uploading

1. Compose your pack JSON (use the format reference as a template).
2. Paste the complete JSON into the textarea.
3. Click Upload Pack.
4. A status message confirms success or shows the validation error.

Updating an existing pack

To publish a new version of an existing pack, upload the same name with a higher version number (e.g., "1.1.0"). The old version remains applied to any projects that used it. Users who apply the pack after the upload will receive the new version.

To replace an existing entry with the same version (e.g., to fix a typo), re-upload with the same name + version. The existing entry is overwritten.

Vertical classification

Set "vertical" to one of the following values to control sidebar filtering on the Starter Packs page:

• "technical": engineering, data platform, and software packs (default)
• "finance": GAAP, IFRS, SEC reporting, audit, and accounting packs

If vertical is omitted, the system will attempt to infer it from the pack name and description using keyword detection. Setting it explicitly is always more reliable.

Performance Dashboard Admin

The Performance Dashboard shows aggregate feedback statistics across all advisory profiles and users. Use it to monitor response quality and understand how well the advisor is calibrated.

KPI strip

Profile Activity chart

Shows how many feedback samples each reasoning profile has accumulated. Profiles with more samples have been selected more often. Low sample counts mean a profile is still building its performance record.

Profile Leaderboard

Ranks profiles by average score. The active profile (currently answering questions) is the highest-scoring profile with at least 10 samples. Profiles below that threshold are on standby, accumulating samples before they can compete.

Performance Trend (7 days)

Daily average score across all profiles for the past week. An upward slope indicates recent responses have been rated more positively. Sharp dips may coincide with complex question types or a profile rotation event.

Feedback Breakdown

Per-profile view showing average score, sample count, acceptance rate, and status (Active or Standby). Use this table to identify which profiles are performing well and which are dragging the average down.

Outcome Breakdown (30 days)

Distribution of feedback outcome types over the last 30 days:

• Accepted: response used as-is (score: 100)
• Rejected: response was not usable (score: 0)

A healthy distribution is acceptance-heavy with minimal rejections.

Knowledge Graph Admin

The Knowledge Graph provides a visual representation of all knowledge items in a project and the relationships between them. It reveals how decisions connect to the rules derived from them, how notes reference decisions, and how items supersede one another.

KPI strip

• Total Nodes: total knowledge items in the project
• Total Edges: total relationships between items
• Max Traversal Depth: deepest chain of relationships found
• Node Types: number of distinct item types present

Filter bar

Refine what you see in the graph using four filter dimensions:

• Project: select which project's graph to display. The graph clears and reloads when you change projects.
• Node Types: toggle visibility of Rule, Decision, Code, and Note nodes individually.
• Max Depth: limit the traversal depth (1, 2, or 3 hops from any node). Lower depth reduces visual noise in densely connected graphs.
• Edge Types: show or hide specific relationship types: applies_to, references, derived_from, supersedes.

Interacting with the graph

• Click a node: opens the Node Detail drawer on the right, showing the item's full content, type badge, metadata, and list of connections.
• Drag nodes: reposition them for clarity.
• Scroll to zoom: zoom in or out on dense graphs.
• Focused mode: when a node is selected, the graph may highlight its neighborhood. Click × Reset to return to the full graph view.

Node colors

Bottom analytics tables

• Node Type Distribution: count and percentage share of each node type
• Edge Type Distribution: count and percentage share of each edge type

Knowledge Item Types: Reference

Choosing the right item type is important. The advisor is aware of types and uses them to weight and present information appropriately. Rules carry authority; decisions carry history; notes carry context.

Rule

A rule is a standing directive: something that is always true or always expected of the team. Rules do not expire unless superseded. Rationale is optional but helps the advisor explain the rule when asked.

Decision

A decision is a choice made at a point in time. Decisions can be revisited; the current decision is the one that stands until a new decision explicitly supersedes it. Rationale is required for all decisions because the value of recording a decision is understanding why it was made, not just what was decided.

Note

A note is flexible context: background information, references to external documents, caveats, or anything that helps the advisor give a richer, more accurate answer without being a rule or decision. Use notes liberally.

Code Context

Code context captures technical patterns, naming conventions, architecture patterns, or implementation snippets relevant to the project. This type is especially useful for grounding the advisor in the team's actual code style so that technical guidance is precise and immediately applicable.

Pack JSON Format: Reference

This reference describes every field in the Starter Pack JSON format. Use this when creating packs for upload via the Admin page.

Top-level fields

Item object fields

Complete example

{
  "name": "snowflake-governance",
  "version": "1.0.0",
  "description": "Core governance rules and architectural decisions for Snowflake data platform implementations. Covers security, cost management, and query performance.",
  "vertical": "technical",
  "items": [
    {
      "item_type": "rule",
      "content": "All Snowflake roles must follow the RBAC hierarchy: SYSADMIN > DATABASE_OWNER > SCHEMA_OWNER > object-level roles. No user should be granted ACCOUNTADMIN for day-to-day operations."
    },
    {
      "item_type": "decision",
      "content": "We use transient tables for all intermediate transformation layers.",
      "rationale": "Transient tables do not contribute to Fail-safe storage costs and our intermediate data is reproducible from source. This reduces monthly storage costs by approximately 40% compared to permanent tables."
    },
    {
      "item_type": "note",
      "content": "Snowflake credits are billed per-second with a 60-second minimum per warehouse start. Auto-suspend should be set to 60 seconds for development warehouses and 5 minutes for production warehouses."
    },
    {
      "item_type": "code_context",
      "content": "Warehouse naming convention: {ENV}_{TEAM}_{SIZE}_WH. Example: PROD_ANALYTICS_MEDIUM_WH. Environment prefixes: DEV, STG, PROD."
    }
  ]
}

Frequently Asked Questions