Semantic Libraries: Knowledge 📚

Broad general knowledge is great, but your agents need to know about your business: company FAQs, product API docs, database schemas, or customer records.

For local 1-7B parameter models, trying to memorize everything or feeding entire files into the prompt window leads to confusion.

Knowledge gives your agents a semantic search library! 📚

Load your files, URLs, or markdown documentation into a local vector database. When a user asks a question, the agent searches the database, finds the most relevant paragraphs, and references them to answer accurately.

🚀 Ingesting Knowledge in 3 Lines

Here is how to boot up a local knowledge base with ChromaDB, insert a web document, and chat with an Ollama-powered agent about it:

1from kern.agent import Agent
2from kern.knowledge.knowledge import Knowledge
3from kern.vectordb.chroma import ChromaDb
4from kern.models.openai import OpenAIChat
5
6# 1. Establish our local vector store 🗄️
7knowledge = Knowledge(
8    vector_db=ChromaDb(collection="my_docs", path="tmp/chromadb"),
9)
10
11# 2. Feed it some content (supports PDFs, text, markdown, web URLs, etc.) 📂
12knowledge.insert(url="https://kern.ndx.rocks/getting-started")
13
14# 3. Create an agent equipped to search this knowledge base
15agent = Agent(
16    model=OpenAIChat(id="llama3.2:3b", base_url="http://localhost:11434/v1"),
17    knowledge=knowledge,
18    search_knowledge=True,  # Let the agent search when needed! 🔍
19)
20
21agent.print_response("How do I install the kern-ai pip package?")

⚙️ The RAG Pipeline: How It Works

Building a knowledge base involves three automated phases:

1. Ingestion (Reading): Kern reads documents from local directories, PDFs, raw text, or web links.
2. Chunking & Embedding (Indexing): Documents are cut into smaller paragraphs (chunks) and converted into vector coordinates (embeddings) representing their semantic meaning.
3. Retrieval (Searching): When the user asks a question, Kern finds the chunks that are mathematically closest in meaning, extracts them, and hands them to the model as context.

🧠 Traditional RAG vs. Agentic RAG

• Traditional RAG: Always searches the database and forces the paragraphs into the prompt, whether the model needs them or not.
• Agentic RAG (Default in Kern) 🤖: The agent decides if it needs to search. If the user says "Hello!", it just responds. If the user says "How does the widget compile?", the agent triggers a search tool, reads the documentation, and answers. This saves tokens and keeps local models focused!

✍️ Agents That Learn Over Time

Knowledge isn't just read-only. You can write custom tools that allow your agent to write its own findings back into the knowledge base:

1def save_new_insight(title: str, content: str) -> str:
2    """Save an important learning or user fact into the knowledge base.
3
4    Args:
5        title (str): The name or category of this insight.
6        content (str): The details of the learning.
7    """
8    knowledge.insert(name=title, text_content=content)
9    return f"Insight saved under: '{title}'"
10
11agent = Agent(
12    model=OpenAIChat(id="llama3.2:3b", base_url="http://localhost:11434/v1"),
13    knowledge=knowledge,
14    search_knowledge=True,
15    tools=[save_new_insight],  # Hand the write tool to the agent! ✍️
16)

Now, your agent can dynamically learn, adapt, and build its own custom database over time!

📚 Explore the Docs