lancedb/docs/web/integrations/ai/agno.mdx

---
title: "Agno"
sidebarTitle: "Agno"
description: "Build a search assistant using the Agno agent framework with LanceDB as the knowledge backend."
---

import {
  PyFrameworksAgnoAgent,
  PyFrameworksAgnoCliChat,
  PyFrameworksAgnoIngestYoutube,
  PyFrameworksAgnoSetup,
} from '/snippets/integrations.mdx';

[Agno](https://docs.agno.com/introduction) is a framework for building agentic AI applications.
It supports LanceDB as a knowledge backend, allowing you to easily ingest and retrieve external content for your agents.

When you pair Agno's `Knowledge` system with LanceDB, you get a clean Agentic RAG setup.
We'll walk through the steps below to build a YouTube transcript-aware Agno assistant that can:
- Ingest a transcript from a YouTube video via the YouTube API
- Store embeddings and metadata in LanceDB
- Retrieve context during responses with hybrid search
- Ask questions about the video content in a CLI chat loop

## Prerequisites {#prerequisites}

Install dependencies:

<CodeGroup>
```bash pip icon="terminal"
pip install -U agno openai lancedb youtube-transcript-api beautifulsoup4
```

```bash uv icon="terminal"
uv add agno openai lancedb youtube-transcript-api beautifulsoup4
```
</CodeGroup>

## Step 1: Configure LanceDB-backed knowledge {#step-1-configure-lancedb-backed-knowledge}

First, you can initialize the core `Knowledge` object that your agent will use for retrieval.
It configures LanceDB as the vector store, enables hybrid search with native LanceDB FTS, and sets the embedding model.

<CodeBlock filename="Python" language="Python" icon="python">
  {PyFrameworksAgnoSetup}
</CodeBlock>

## Step 2: Fetch and ingest the YouTube transcript {#step-2-fetch-and-ingest-the-youtube-transcript}

Next, extract a YouTube video ID, fetch the full transcript, and flatten it into text for indexing.
The snippet shown below then inserts that transcript text into the Agno knowledge base, which writes vectors and metadata to LanceDB.

<CodeBlock filename="Python" language="Python" icon="python">
  {PyFrameworksAgnoIngestYoutube}
</CodeBlock>

<Info>
This path explicitly fetches the transcript first, then inserts transcript text into LanceDB through Agno.
</Info>

## Step 3: Create an Agno agent with knowledge search {#step-3-create-an-agno-agent-with-knowledge-search}

The next step is to construct an Agno `Agent` and attach the knowledge base you just populated.
With `search_knowledge=True`, the agent performs retrieval before answering, so responses stay grounded in transcript context.

In Agno, retrieval is exposed as a tool call that the model can invoke at runtime.
When `search_knowledge=True`, Agno makes a knowledge-search tool (shown in output as `search_knowledge_base(...)`) available to the model; the model decides when to call it, Agno executes the tool, and the returned context is fed back into the final answer.

<CodeBlock filename="Python" language="Python" icon="python">
  {PyFrameworksAgnoAgent}
</CodeBlock>

## Step 4: Start a CLI chat loop {#step-4-start-a-cli-chat-loop}

You can now ask an initial question and then start an interactive loop for follow-up queries.
Each prompt runs through the same retrieval pipeline, so you can iteratively inspect what the transcript contains.

<CodeBlock filename="Python" language="Python" icon="python">
  {PyFrameworksAgnoCliChat}
</CodeBlock>

<Info>
Want local-first inference? Replace OpenAI model/embedder classes with Agno's Ollama providers. See Agno's Ollama knowledge examples: [docs.agno.com/examples/models/ollama/chat/knowledge](https://docs.agno.com/examples/models/ollama/chat/knowledge).
</Info>

### Question 1 {#question-1}

The following question is asked in the CLI chat loop:
```
┏━ Message ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                  ┃
┃ Q: What kinds of data can LanceDB handle?                                        ┃
┃                                                                                  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━ Tool Calls ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                  ┃
┃ • search_knowledge_base(query=What kinds of data can LanceDB handle?)            ┃
┃ • search_knowledge_base(query=LanceDB images audio video handle kinds of data    ┃
┃ can handle 'LanceDB can handle' 'kinds of data' 'images audio video' transcript) ┃
┃                                                                                  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━ Response (19.1s) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                  ┃
┃                                                                                  ┃
┃  • Images, audio, video — i.e., multimodal AI data and “all manners of things    ┃
┃    you don't put into traditional databases” (per the transcript).               ┃
┃                                                                                  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
```

We get the response based on the transcript's contents as expected.

### Question 2 {#question-2}

Let's ask a more specific question about the CEO of LanceDB, which is also in the transcript:

```
You: What is the name of the CEO of LanceDB? 
INFO Found 10 documents                                                             
┏━ Message ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                  ┃
┃ What is the name of the CEO of LanceDB?                                          ┃
┃                                                                                  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━ Tool Calls ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                  ┃
┃ • search_knowledge_base(query=CEO of LanceDB)                                    ┃
┃                                                                                  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━ Response (16.7s) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                  ┃
┃                                                                                  ┃
┃  • According to the retrieved YouTube transcript/title, the CEO of LanceDB is    ┃
┃    Chang She.                                                                    ┃
┃                                                                                  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
```

We get the response based on the transcript's contents and title as expected.

## Why this works well {#why-this-works-well}

To start, LanceDB Embedded can run from a local directory, so transcript data can stay on your machine.

- You do not need to maintain a separate transcript parser in your application code.
- You do not need to hand-roll chunking and retrieval orchestration across multiple modules.
- One explicit Agno `Knowledge` object, backed by LanceDB, defines both ingestion and search behavior in one place.
- Fewer moving parts means the tutorial stays readable and the same pattern is easier to carry into production code.

As your application needs grow, you can migrate to LanceDB [Enterprise](/enterprise) for
convenience features like automatic compaction and reindexing and the ability to scale to
really large datasets.