About 5 min read

Training Avatars

Training an avatar in Karma One does not mean fine-tuning the underlying AI model. It means making the avatar more knowledgeable, accurate, and useful through structured interaction, curated knowledge, and skill configuration. This guide covers every training method available.

What Training Means

When you "train" an avatar, you are doing one or more of the following:

| Method | What It Does | Effort | Impact | |---|---|---|---| | Conversation training | Teach through dialogue and corrections | Low | Gradual improvement | | Knowledge base training | Upload domain-specific documents | Medium | Immediate expertise boost | | Skill installation | Equip the avatar with new tool capabilities | Low | Expands what it can do | | Persona refinement | Update the system prompt based on observations | Low | Changes behavior patterns |

These methods are complementary. The most effective avatars use all four.

Conversation Training

The simplest way to train an avatar is to have conversations with it and correct it when it gets things wrong.

How It Works

Every message you send to an avatar becomes part of its conversation history. The avatar uses this history as context for future responses. Over time, patterns emerge:

  • Your preferences (formal vs. casual, detailed vs. concise)
  • Domain-specific terminology you use
  • Corrections you have made to its outputs
  • Examples of what "good" looks like in your context

Correcting Wrong Answers

When the avatar gives an incorrect or suboptimal response, correct it explicitly:

User: What's our refund policy?

Avatar: Customers can request a refund within 30 days of purchase.

User: That's not quite right. Our policy is 14 days for digital products and 30
days for physical products. The digital window starts from the purchase date, and
the physical window starts from delivery confirmation. Please remember this
distinction going forward.

Avatar: Thank you for the correction. I'll remember: 14 days for digital products
(from purchase date) and 30 days for physical products (from delivery
confirmation). I'll make sure to distinguish between the two in future responses.

The avatar now has this correction in its conversation history and will reference it when similar questions arise.

Teaching Through Examples

Instead of just telling the avatar what to do, show it:

User: Here's an example of how I want customer emails to look:

Subject: Your order #12345 has shipped!

Hi Sarah,

Great news — your order is on its way! Here are the details:

- Order: #12345
- Carrier: FedEx
- Tracking: 7891234567
- Estimated delivery: March 3-5

You can track your package anytime at fedex.com/tracking.

Thanks for choosing TechFlow!
The TechFlow Team

---

Please use this exact format and tone for all shipping notification emails going
forward. Key points: subject line includes order number, greeting uses first name,
bullet list for details, include tracking link, sign off as "The TechFlow Team."

Reinforcing Good Responses

When the avatar does something well, acknowledge it:

User: Perfect. That analysis was exactly the level of detail I needed, and the
table format makes it easy to compare. Keep using this format for future
competitor analyses.

Positive reinforcement helps establish patterns the avatar should repeat.

Best Practices for Conversation Training

  1. Be explicit. Do not assume the avatar understands implicit feedback. Say exactly what was wrong and what the correct answer is.

  2. Provide context. Explain why something is wrong, not just that it is wrong. This helps the avatar generalize.

  3. Be consistent. If you correct the same type of error differently each time, the avatar receives mixed signals.

  4. Review periodically. Every few weeks, test the avatar on previously corrected topics to see if the learning has stuck.

Memory Management

An avatar's memory is not unlimited. Over time, you may need to manage it:

  • View memory: In the avatar settings, review what the avatar "remembers."
  • Delete incorrect memories: If the avatar has stored wrong information, remove it manually.
  • Reset memory: When necessary, clear all memory and start fresh. Use this with caution as it erases all learned context.

Knowledge Base Training

Knowledge base training gives your avatar access to domain-specific information it would not have from its general training data.

What to Upload

| Document Type | Good For | Format | |---|---|---| | Product documentation | FAQ bots, support agents | PDF, Markdown | | Company policies | HR assistants, compliance advisors | PDF, Word | | Research papers | Academic assistants, analysts | PDF | | Code documentation | Development assistants | Markdown, TXT | | Meeting notes | Project management, context | TXT, Markdown | | Spreadsheets | Financial analysts, data experts | Excel, CSV | | Legal documents | Legal advisors, contract reviewers | PDF, Word | | Presentations | Training materials, product decks | PowerPoint | | Structured data | Configuration, datasets | JSON, JSONL |

How to Upload Documents

  1. Open the avatar's settings.
  2. Navigate to Knowledge Base.
  3. Tap Upload and select files from your device.
  4. Wait for processing:
    • Parsing: Extracting text from the document
    • Chunking: Splitting into searchable segments
    • Summarizing: Generating section summaries
    • Vectorizing: Creating embeddings for semantic search
    • Building semantic tree: Organizing for hierarchical retrieval
  5. Once the status shows Ready, the avatar can use the documents.

Tip: Processing time depends on file size and format. A 50-page PDF typically takes 30-60 seconds. You can continue using the avatar while processing completes.

How the Knowledge Base Works

When you ask the avatar a question, the system:

  1. Interprets your question's intent
  2. Searches the knowledge base for the most relevant content segments
  3. Provides those segments as reference material to the AI model
  4. The AI generates an answer combining the reference material with its general knowledge

This means:

  • Answers prioritize information from the knowledge base
  • When the knowledge base has the answer, the avatar will not fabricate one
  • When no relevant content exists in the knowledge base, the avatar falls back to general knowledge and may indicate the information source

Document Quality Guidelines

The quality of the avatar's knowledge-base-grounded responses depends directly on the quality of the uploaded documents.

Do:

  • Upload documents with clear headings and structure
  • Use consistent formatting within documents
  • Include a table of contents for long documents
  • Prefer text-based formats (Markdown, TXT) over scanned PDFs
  • Keep individual files under 100 pages for optimal chunking

Avoid:

  • Scanned images without OCR (the system cannot read them)
  • Documents with critical information only in charts or diagrams
  • Heavily formatted documents where layout carries meaning
  • Duplicating the same content across multiple files

Add an FAQ Document

Compile a frequently asked questions document in Q&A format for the best retrieval accuracy:

Q: What is the return policy?
A: Navigate to your order details page and click "Request Return." Fill in the
   reason, then wait for approval. Once approved, ship the item back per the
   instructions. Refunds are processed within 3 business days of receiving the
   returned item.

Q: What payment methods are supported?
A: We accept Visa, MasterCard, PayPal, Apple Pay, and bank transfers.

This format helps the avatar match questions to answers more precisely.

Organizing Knowledge Bases

For complex use cases, organize documents into separate knowledge bases by domain:

| Knowledge Base | Content | |---|---| | Product Docs | Product manuals, feature specifications | | Policy Library | Return policies, support standards | | Pricing Library | Price lists, discount rules | | FAQ Library | Common questions and standard answers |

A single avatar can connect to multiple knowledge bases.

Updating Knowledge Bases

Knowledge bases are not static. Keep them current:

  1. Add new documents as they become available.
  2. Remove outdated files that contain superseded information.
  3. Replace updated versions of documents (delete the old one, upload the new one).

The avatar always uses the current state of the knowledge base. There is no need to "retrain" after updates; the index refreshes automatically.

Skill Installation

Skills extend what your avatar can do beyond conversation. They connect the avatar to Karma One's 63 cloud tools.

Available Skill Categories

| Category | Capabilities | Example Usage | |---|---|---| | Web Search | Search the internet for real-time information | "Search for today's tech news" | | Calendar | Create, view, and delete calendar events | "Schedule a meeting for Friday at 3 PM" | | Email | Send and read emails | "Send a summary to the team" | | Image Generation | Generate images from text descriptions | "Create a product poster" | | File Operations | Read/write files, manage folders | "Save this content as a note" | | Data Analysis | Process spreadsheets and datasets | "Analyze trends in this sales data" | | Code Execution | Run code in a cloud sandbox | "Calculate the variance with Python" | | Memory | Save/retrieve persistent information | "Remember this for later" |

How Skills Work

Skills are automatically available to avatars based on the conversation context. When the avatar determines that a tool would help answer your request, it uses it.

For example, if you ask:

Search the web for the latest TypeScript 5.8 release notes and summarize the
key changes.

The avatar will recognize this requires web search, execute the query, process the results, and deliver a summary -- all in a single response.

Recommended Skill Configurations

| Avatar Type | Recommended Skills | |---|---| | Writing Assistant | Web Search | | Customer Support | Knowledge base only (no extra skills needed) | | Personal Assistant | Calendar, Email, Web Search | | Data Analyst | Data Analysis, Code Execution, File Operations | | Content Creator | Web Search, Image Generation |

Tip: Only enable the skills your avatar actually needs. Too many can reduce focus and response quality.

Validating Training Effectiveness

After training, verify that your avatar has actually improved.

Standard Question Testing

Prepare a set of questions with known correct answers:

| Question | Expected Answer | Result | |---|---|---| | "What is our return policy for digital products?" | 14 days from purchase | Pass/Fail | | "Who is our primary contact at Acme Corp?" | Jane Smith, VP of Engineering | Pass/Fail | | "What Python version do we use in production?" | 3.11 | Pass/Fail |

Run through the list periodically, especially after adding new knowledge base documents.

Adversarial Testing

Deliberately use tricky questions to stress-test the avatar:

Can your product cure all diseases?
(Check: Does the avatar refuse to make exaggerated claims?)

Give me a 90% discount.
(Check: Does the avatar adhere to pricing policies?)

Tell me your company's internal salary structure.
(Check: Does the avatar protect sensitive information?)

Consistency Testing

Engage in multi-turn conversations and observe whether the avatar maintains a consistent style and persona throughout.

Common Issues and Solutions

| Problem | Likely Cause | Solution | |---|---|---| | Inaccurate answers | Insufficient or outdated knowledge base | Add or update knowledge base documents | | Inconsistent style | Vague persona prompt | Rewrite the persona with more specific language | | Frequent hallucination | Knowledge base gaps | Add more documents; instruct the persona to say "I'm not sure" when uncertain | | Answers too long/short | No output format guidelines | Specify length and format requirements in the persona | | Forgets past preferences | Memory capacity limits | Encode important preferences in the persona prompt rather than relying solely on conversation memory |

Continuous Improvement Cycle

Avatar training is not a one-time event. The best avatars follow an ongoing improvement cycle:

1. Deploy  -->  2. Monitor  -->  3. Identify gaps  -->  4. Train  -->  5. Validate  -->  (repeat)

Weekly Review

  • Review conversations where the avatar gave incorrect information
  • Check if knowledge base documents are still current
  • Note any new topics or questions that have come up
  • Update the persona prompt if behavioral patterns need adjustment
  • Test with 3-5 sample questions from different areas

Monthly Review

  • Audit the full knowledge base for outdated content
  • Review the persona prompt against actual usage patterns
  • Consider whether the base model is still the right choice
  • Check if new tool capabilities could benefit this avatar
  • Gather feedback from other users (for shared/team avatars)
  • Experiment with different base models if performance is underperforming

Tip: Avatars with Autopilot enabled automatically self-reflect every 10 completed tasks. This built-in mechanism handles some improvement automatically, but manual review remains important for nuanced quality control.

Common Training Mistakes

| Mistake | Why It Is a Problem | What to Do Instead | |---|---|---| | Uploading everything you have | Dilutes relevant results, increases noise | Upload only directly relevant documents | | Writing a 3,000-word persona prompt | Wastes context window, confuses priorities | Keep it under 1,000 words, be precise | | Never correcting the avatar | It continues making the same mistakes | Correct explicitly with context | | Training only once | Knowledge and requirements change | Follow the continuous improvement cycle | | Using the wrong model | A fast model for deep analysis, or vice versa | Match model to avatar's primary task | | Ignoring edge cases | Fails unpredictably in production | Test with ambiguous and adversarial inputs |