The Knowledge Base (KB) is your agent's knowledge store. Instead of the AI answering only from its general knowledge, the KB allows the agent to use your specific data — documentation, FAQ pages, blog posts, manuals.

Add source (URL / text / file)
    ↓
Scraping / parsing content
    ↓
Chunking (splitting text into ~400 token pieces)
    ↓
Embedding (converting each chunk to a 1536-dimension vector)
    ↓
Saving to PostgreSQL (pgvector)

When a user asks a question, the agent:

1. Converts the question into an embedding vector
2. Finds the most similar chunks (semantic search + BM25 keyword search)
3. Merges results (Reciprocal Rank Fusion)
4. Optionally re-ranks using an LLM
5. Passes the context to the AI Response node

Scrape content from any web page. Agent Studio uses the Cheerio parser which automatically removes navigation, footer, script, and style tags — only useful text remains.

Directly enter text. Useful for internal documents, FAQ answers, or anything not available as a URL.

Upload PDF or DOCX files (max 10 MB). Supported formats:

• PDF — parsed using the pdf-parse library
• DOCX — parsed using the mammoth library

1. Go to the agent's Knowledge page
2. Click "Add Source"
3. Select the "URL" tab
4. Enter the full URL (with https://)
5. Click "Add"

There is no strict limit, but here are guidelines:

It's better to have 10 URLs with clean, relevant content than 100 URLs with noise. Every URL you add goes through chunking and indexing — irrelevant chunks can reduce answer quality.

On the Knowledge page, each source displays a status:

Next to each source, the number of chunks is displayed (e.g. "24 chunks"). If a source has 0 chunks but its status is READY, the page probably doesn't have enough text.

The best way to verify is to test the search:

1. On the Knowledge page, use the Search functionality
2. Enter a query related to the source content
3. Check if results return relevant chunks
4. Pay attention to the score — higher score means greater relevance

Pages with clear headings, paragraphs, and lists produce better chunks than pages with lots of navigation and ads.

Bad:  https://company.com
Good: https://company.com/docs/installation
Good: https://company.com/faq/payment

If you have information that's not on the web (business hours, prices, policies), add it as a Text source. Format it however you want — this produces the cleanest chunks.

If users ask questions in one language but KB content is in another, search quality will be lower. Try to have KB content in the same language as the expected questions.

The System Prompt in the AI Response node should tell the agent:

• Use only information from the context
• If there's no answer in the context, tell the user
• Respond in the user's language

Example:

You are an assistant for Company X. Answer only based on the provided context.
If the context does not contain the requested information, tell the user you 
don't have that information and direct them to support@company.com.
Be concise — respond in 2-3 sentences when possible.

1. Add sources
2. Test with real questions
3. If answers aren't good enough:
   • Add more specific URLs
   • Add Text sources with missing information
   • Improve the System Prompt
4. Repeat until satisfied

If a source is not useful or contains outdated information:

1. On the Knowledge page, find the source
2. Click the Delete button
3. The source and all its chunks are deleted from the database

Deletion is permanent — you'll need to re-add the URL if you want it back.

Agent Studio uses several techniques beyond basic vector search to deliver more accurate and complete answers.

When a search finds a relevant chunk, the system automatically expands the context by including neighboring chunks from the same source. This means even if the answer spans multiple chunks, the AI sees the full picture.

For example, if chunk #5 of a page is the best match, the system will also retrieve chunks #4 and #6 to provide surrounding context. This prevents answers from being cut off mid-thought.

Results with a relevance score below 0.25 are filtered out automatically. This prevents low-quality chunks from polluting the context and confusing the AI. If none of the chunks meet the threshold, the agent receives empty context — which is better than irrelevant context.

The number of results retrieved adapts to the query complexity:

Short queries are usually specific (e.g. "pricing"), so fewer results are needed. Longer queries benefit from more context.

Search combines two methods with weighted scoring:

• Semantic search (70%) — finds content with similar meaning, even if different words are used
• BM25 keyword search (30%) — finds exact keyword matches, good for technical terms and proper nouns

This weighted approach means semantic similarity drives most results, but exact matches still get a boost.

Retrieved chunks are capped at 4000 tokens total. This prevents overloading the AI's context window with too much KB content, leaving room for conversation history and system prompt instructions.

• Chunk size: ~400 tokens with 20% overlap between chunks
• Embedding model: OpenAI text-embedding-3-small (1536 dimensions)
• Search: Hybrid (semantic cosine similarity + BM25 keyword search)
• Ranking: Reciprocal Rank Fusion (70/30 semantic/keyword) + optional LLM re-ranking
• Parent retrieval: Automatic expansion to neighboring chunks within 4000 token budget
• Similarity threshold: 0.25 minimum relevance score
• Storage: PostgreSQL with pgvector extension
• Max upload: 10 MB per file (PDF/DOCX)