Orq.ai Documentation - AI Gateway & LLM Collaboration Platform

Creating a Knowledge Base

Use the + button in a chosen Project and select Knowledge Base > Internal. Press Create Knowledge, the following modal will appear:

You can only create a Knowledge Base once you have activated an embedding model within your Model Garden

With the Knowledge Bases API, it is now possible to create Knowledge Bases with code, to learn more, see Creating a Knowledge Base Programmatically.

Adding a source

You are then taken to the source management page. A source represents a document that is loaded within your Knowledge Base. This document’s information will then be used when referencing and querying the Knowledge Base. Documents needs to be loaded ahead of time so that they can be parsed and cut into chunks. Language models will then use the loaded information as source for answering user queries. To load a new source, select the Add Source button. Here you can add any document of the following format: TXT, PDF, DOCX, CSV, XML.

While you can add any number of sources to a Knowledge Base, A single source document must be of a maximum of 10MB.

Once you have selected files from your disk, you will be able to configure how the file is parsed and indexed within the Knowledge Base.

Chunk Settings and Strategies

Chunks are portions of a source document loaded within a Knowledge Base. When adding a new source to a Knowledge Base, you can decide how this source’s information will be chunked. Larger chunks will hold more relevant information but will imply more token use when sent to a model, impacting the generation cost.

Token

Splits text into chunks based on token count. Best for ensuring chunks fit within LLM context windows and maintaining consistent chunk sizes for embedding models.

Parameter	Description	Default
`chunk_size`	Maximum tokens per chunk	512
`chunk_overlap`	Number of tokens to overlap between chunks	0

Sentence

Splits text at sentence boundaries while respecting token limits. Ideal for maintaining semantic coherence and readability.

Parameter	Description	Default
`chunk_size`	Maximum tokens per chunk	512
`chunk_overlap`	Number of overlapping tokens between chunks	0
`min_sentences_per_chunk`	Minimum number of sentences per chunk	1

Recursive

Recursively splits text using a hierarchy of separators (paragraphs, sentences, words). Versatile general-purpose chunker that preserves document structure.

Parameter	Description	Default
`chunk_size`	Maximum tokens per chunk	512
`separators`	Hierarchy of separators to use	`["\n\n", "\n", " ", ""]`
`min_characters_per_chunk`	Minimum characters allowed per chunk	24

Semantic

Groups semantically similar sentences using embeddings. Excellent for maintaining topic coherence and context within chunks.

Parameter	Description	Default
`chunk_size`	Maximum tokens per chunk	512
`embedding_model`	Embedding model for similarity (required)	-
`dimensions`	Number of dimensions for embedding output	-
`threshold`	Similarity threshold (0-1) or “auto"	"auto”
`mode`	Chunking mode: “window” or “sentence"	"window”
`similarity_window`	Window size for similarity comparison	1

Agentic

AI-powered intelligent chunking that uses an LLM to determine optimal split points. Best for complex documents requiring intelligent segmentation.

Parameter	Description	Default
`model`	LLM model to use for chunking (required)	-
`chunk_size`	Maximum tokens per chunk	1024
`candidate_size`	Size of candidate splits for LLM evaluation	128
`min_characters_per_chunk`	Minimum characters allowed per chunk	24

Fast

High-performance SIMD-optimized byte-level chunking. Best for large files (>1MB) where speed and memory efficiency are critical. 2x faster and 3x less memory than token-based chunking.

Parameter	Description	Default
`target_size`	Target chunk size in bytes	4096
`delimiters`	Single-byte delimiters to split on (e.g., `"\n.?!"`)	`"\n.?"`
`pattern`	Multi-byte pattern for splitting (e.g., `"▁"` for SentencePiece)	-
`prefix`	Attach delimiter to start of next chunk	false
`consecutive`	Split at START of consecutive delimiter runs	false
`forward_fallback`	Search forward if no delimiter found backward	false

When to use Fast: Large files (>1MB), high-throughput ingestion, memory-constrained environments.When NOT to use Fast: When you need precise token counts for embedding models, small documents where speed isn’t critical, or when semantic boundaries matter more than byte boundaries.

Strategy Selection Guide

Use Case	Recommended Strategy
Large files (>1MB)	Fast - 2x faster, 3x less memory
RAG with precise tokens	Token or Recursive
Semantic search	Semantic
Complex document understanding	Agentic
General purpose	Recursive

UI Configuration

Default

Automatically set chunk and preprocessing rules. Unfamiliar users are recommended to select this.

Advanced

Maximum Chunk LengthDefines the maximum size of each chunk. The bigger the size, the more information they contain.Chunk OverlapDefines the number of characters overlapping neighboring chunks. The higher the value, the more chunks will contain redundant information from one another, but the more likely relevant information will be sent back to models.

Data Cleanup

You can choose to modify the data loaded within your sources, this can be great to clean the chunks or anonymize data. To activate each cleanup, simply toggle on the option within the data cleanup panel.

Summary and Cost Estimation

Once your document has been processed, the following summary will be displayed:

Retrieval Settings

You can configure these options on the Knowledge Settings page. Each option will yield different results, depending on your needs.

Search Methods

Vector Search

Vector search is the fastest method of searching through a database built with your Knowledge Sources. Here, our systems take the user query and look for the text segments most similar to their vector representations.The search will return the preprocessed chunks from the sources most similar and relevant to the user’s query.

Keyword Search

Keyword Search is a different method for retrieving relevant results within a Knowledge Base. In this method, the entire content is indexed, and the system searches for segments containing the words from the user’s query.

Hybrid Search

Hybrid search uses both the previously mentioned Vector & Keyword searches, then combining results and returning the most relevant chunks to the model.

Search Parameters

All previous search types can be configured with the following parameters:

Chunk limit

This parameter sets the number of chunks most similar to the user’s questions.

Threshold

This controls the relevance of the results on a scale from 0 to 1. Results scoring below threshold will be excluded from retrieval.The closer to 1, the more relevant and narrow the results will be.

Setting a too high threshold can yield little to no result to a search.

Reranking

Reranking invokes a model that analyzes your initial query and the result fetched by the Knowledge Base search. This model then scores the similarity of the chunks returned with the user query, then scores and ranks the chunks accordingly. This ensures the results is the most relevant for your query.

To use reranking within your Knowledge Base, you must enable at least one Reranking model within your Model Garden.

Knowledge Settings

By choosing the Knowledge Settings button, you can configure the following settings.

Embedding Models

Here, you can configure which llm model to use to query the Knowledge Base. Your configuration here is similar to any model configuration within Playground, Experiment, Deployment, and Agent and includes the usual parameters

Agentic RAG

Settings****Agentic RAG

The agent performs two main actions:

Document Grading, which ensures relevant chunks are retrieved.
Query Refinement, improving the query if needed.

Example

See the screenshot below on how the input query gets refined. Input query: is my suitcase too big? is reformulated to luggage size requirements and restrictions for carry-on and checked baggage

Connecting an External Knowledge Base

To connect to an external Knowledge Base, choose the + button on the desired Project.

The following modal opens to configure an external knowledge base.

Configuration

Field	Description	Example
Key	Unique Identifier, alphanumeric with hyphens/underscore	`external_kb`
Description	Description	`External Knowledge Base`
Name	Display Name	`External Knowledge Base Name`
API URL	URL to search knowledge base, must be HTTPS	`https://api.example.org/search`
API Key	Authentication API key to the previously API URL. Orq will use Bearer Authentication Header to call your API.	`<API_KEY>`

orq.ai will include the API Key in the Authorization: Bearer <API_KEY> header when calling your endpoint.

API keys are encrypted using workspace-specific keys (AES-256-GCM)

Select Connect to finalize connecting your external Knowledge Base.

API Payloads

Here are example payloads for request and response expected from your API.

Request Payload

{
  "query": "<string>",
  "top_k": 50,
  "threshold": 0.5,
  "filter_by": {},
  "search_options": {
    "include_vectors": true,
    "include_metadata": true,
    "include_scores": true
  },
  "rerank_config": {
    "model": "cohere/rerank-multilingual-v3.0",
    "threshold": 0,
    "top_k": 10
  }
}

Response Payload

{
  "matches": [
    {
      "id": "<string>",
      "text": "<string>",
      "vector": [
        123
      ],
      "metadata": {},
      "scores": {
        "rerank_score": 123,
        "search_score": 123
      }
    }
  ]
}

The API must respond like a standard Knowledge Base search, to learn more about the expected payload, see our Search API.

Example Implementation for an External API

We’ve created example implementation for External Knowledge Base API.

Python Implementation

An Example Python Server for External Knowledge Base

Get the Code

Clone the Python example Server

Install Dependencies

# Install dependencies
pip install -r requirements.txt

Run the Server

# Run the server
uvicorn main:app --reload

Test the API

The API is running at http://localhost:8000Dynamic Documentation will be running at http://localhost:8000/docs

Node.js Implementation

An Example Node Server for External Knowledge Base

Get the Code

Clone the Node example Server

Install Dependencies

# Install dependencies
npm install

Run the Server

# Run the server
npm run dev

Test the API

The API is running at http://localhost:8000Dynamic Documentation will be running at http://localhost:8000/doc

Integrating Vector Database Providers

We support providers like Weaviate and Pinecone, as both platforms expose REST APIs that conform to the expected payload format documented above. Integration Examples

Show Weaviate

Configuration in Orq.ai:

API URL: https://your-cluster.weaviate.cloud/v1/graphql
API Key: Your Weaviate API key

Show Pinecone

Configuration in Orq.ai:

API URL: https://$INDEX_HOST/records/namespaces/$NAMESPACE/search
API Key: Your Pinecone API key

Common Errors and Troubleshooting

Scenario	Error Message
HTTP instead of HTTPS	”External knowledge base URL must use HTTPS protocol”
Local/private IP	”External knowledge base URL cannot point to local network”
API unreachable	”Failed to verify external knowledge base connectivity”
API timeout (>50s)	“External API request timed out”

Problem: Cannot connect to external API

Verify your API endpoint is publicly accessible via HTTPS
Check your API logs for incoming requests from orq.ai IP addresses
Verify your firewall/security groups allow inbound HTTPS traffic

Problem: API key authentication failing

Verify the API key is correct and has not expired
Check that your API expects Bearer authentication in the Authorization header
Confirm your API key has the necessary permissions to perform searches

Problem: No results returned or poor quality results

Verify your API returns the expected response format (see Response Payload above)
Check that scores.search_score values are between 0 and 1
Test with different threshold values (lower threshold = more results)
If using reranking, ensure both search_score and rerank_score are provided
Verify your external vector database has sufficient indexed documents

Problem: Slow response times

Monitor your external API response times
Consider implementing caching for frequently searched queries
Optimize your vector database indexes
Check if your external API is rate limiting requests

Configuring your External Knowledge Base

Datasource configuration is not accessible within External Knowledge base, as data is hosted outside of Orq.ai.

The available configurations are:

Agentic RAG, to handle knowledge base search refinement with an agent, through orq.ai’s system.
Search retrieval parameters, Chunk Limit, Search Threshold
Reranking Models

For detailed configuration options, scroll to the Knowledge Settings section above to see all available features for both internal and external Knowledge Bases.

Your External Knowledge Base is connected:

Use it just like any other Knowledge Base, to learn more see Using a Knowledge Base in a Prompt.
Your knowledge base can also be used with Agents, see using Knowledge Bases with Agent.
Your API will be called at runtime when the model needs to perform a search.

Retrieval Logs

When Using a Knowledge Base in a Prompt within Playground, Experiment, Deployment, or Agent, logs are generated and will transparently contain details of how Knowledge Bases were accessed. To find logs, head to the Logs tabs within the module you’re in, then select one log to open the detail panel for one log entry. The following panel will open.

On the right side of the screen you have access to the Retrievals section that details the Knowledge Base used and how it was queried. The Query can be seen, which was used to retrieve the relevant chunks from the Knowledge Base. The Documents show the retrieved chunks, ordered by relevance score.

User Message Augmentation

On the left side of the panel, you can see how the Knowledge Base variable is modified with the retrieval results in blue. The blue parts of the messages are the retrieval results injected into the user message. They will then be used as data with which the model can respond to the user query.

Using the blue text, you can verify that the query is correct and that the expected chunks are loaded into the message.

Getting Started

Reference

Organization

Create Knowledge Base | Internal RAG Setup

Creating a Knowledge Base

Adding a source

Chunk Settings and Strategies

Data Cleanup

Summary and Cost Estimation

Retrieval Settings

Search Methods

Search Parameters

Reranking

Knowledge Settings

Embedding Models

Agentic RAG

Example

Connecting an External Knowledge Base

API Payloads

Example Implementation for an External API

Integrating Vector Database Providers

Common Errors and Troubleshooting

Configuring your External Knowledge Base

Retrieval Logs

User Message Augmentation

Getting Started

Reference

Organization

​Creating a Knowledge Base

​Adding a source

​Chunk Settings and Strategies

​Data Cleanup

​Summary and Cost Estimation

​Retrieval Settings

​Search Methods

​Search Parameters

​Reranking

​Knowledge Settings

​Embedding Models

​Agentic RAG

Example

​Connecting an External Knowledge Base

​API Payloads

​Example Implementation for an External API

​Integrating Vector Database Providers

​Common Errors and Troubleshooting

​Configuring your External Knowledge Base

​Retrieval Logs

​User Message Augmentation

Creating a Knowledge Base

Adding a source

Chunk Settings and Strategies

Data Cleanup

Summary and Cost Estimation

Retrieval Settings

Search Methods

Search Parameters

Reranking

Knowledge Settings

Embedding Models

Agentic RAG

Connecting an External Knowledge Base

API Payloads

Example Implementation for an External API

Integrating Vector Database Providers

Common Errors and Troubleshooting

Configuring your External Knowledge Base

Retrieval Logs

User Message Augmentation