Building AI That Knows Your Data: Inside GraphN Knowledgebases

Turn documents, images, and videos into searchable context for grounded AI agents and workflows.

Published: August 28, 2026
Reading time: 11 min

Large language models are powerful, but they do not automatically understand your organization's documents, images, videos, policies, or institutional knowledge.

GraphN Knowledgebases turn that private content into searchable, reusable context for agents and workflows. They provide the retrieval layer behind document question answering, multimodal search, research assistants, voice agents, structured extraction, and other grounded AI applications.

Unlike a traditional document store, a Knowledgebase understands the meaning of its contents. A workflow can search for relevant information even when the user's wording does not exactly match the source.

Why this is not just document search

Traditional search looks for the same words. Vector search looks for the same meaning.

A keyword engine can find "database access policy" if those words appear in the document. It will often miss "How do I get production Postgres credentials?" even when the answer is sitting in that same policy. A Knowledgebase is built for the second case.

Documents become numbers that capture meaning

When you ingest a file, GraphN does not store it as one searchable blob of text.

It:

Splits the content into overlapping chunks.
Sends each chunk through an embedding model.
Stores the result as a vector—a list of numbers that represents the meaning of that chunk.

A text chunk, an image, and a video clip all go through the same idea: turn the content into a point in a high-dimensional space.

Chunks with similar meaning end up close together. Unrelated chunks end up far apart.

text

"reset my password"          ●
"I forgot my login"          ●     close together
"change account credentials" ●

"quarterly revenue forecast"                     ●   far away

That geometry is the core idea. Search becomes: embed the question, then find nearby points.

Similarity is measured, not guessed

GraphN stores those vectors in Qdrant and compares them with cosine similarity. Two items are related if the angle between their vectors is small, not because they share a substring.

That allows queries and results with almost no overlapping words to match:

Scroll horizontally to compare

Query	Retrieved content
"How do we onboard a contractor?"	A handbook section titled "External workforce access"
"Show the diagram of private networking"	An architecture image with no caption matching those words
"Where does she explain failover?"	A 12-second clip from a 90-minute training video

Keyword search would need the exact phrasing, a synonym list, or a human to tag the file first.

Why chunks matter

If you embedded an entire 80-page PDF as one vector, a question about one paragraph would be diluted by everything else in the document.

Chunking keeps retrieval precise:

A policy handbook becomes many small passages.
A long video can become many time-bounded clips.
An image stays one item because the whole image is the unit of meaning.

The workflow retrieves the relevant pieces, not simply "this giant file might be related."

That is also why a Knowledgebase is different from "upload a document and ask the model about it." Direct document search either:

dumps the whole file into the prompt, which is slow, expensive, and constrained by context limits; or
searches for keywords, which is brittle.

Vector retrieval pulls only the passages that are close to the question, then gives those passages to the agent.

Reranking is a second opinion

The first search pass is fast and broad: find the nearest neighbors in vector space.

GraphN can then rerank those candidates with a stronger model:

Text Knowledgebases use a text cross-encoder.
Multimodal Knowledgebases use a vision-language reranker.

Think of vector search as "find the neighborhood," and reranking as "walk the block and pick the right house."

The same idea works across text, images, and video

A conventional document index is built around words. A GraphN Knowledgebase is built around embeddings.

That is why one service can support manuals, contracts, screenshots, diagrams, and training videos with the same retrieval question: what is closest in meaning to this query?

Scroll horizontally to compare

	Direct document or keyword search	GraphN Knowledgebase
Matches	Shared words, filenames, and tags	Shared meaning
Unit of retrieval	Usually the whole file	Chunks, images, or video segments
Synonyms	Need extra rules	Emerge naturally in embedding space
Images and video	Need captions or metadata	Embedded as content
Agent use	Put the file into context, or miss it	Retrieve only the relevant evidence

The language model still does the reasoning. The Knowledgebase's job is to find the right evidence first—even when the user never uses the same words as the source.

One knowledge layer, multiple modalities

GraphN supports Knowledgebases built for different kinds of content. The retrieval idea stays the same; the embedding model and unit of meaning change.

Text and documents

The default text model uses BGE-M3 embeddings with a dedicated reranker. It is well suited to:

Internal documentation
Policies and procedures
Product specifications
Research material
Contracts and legal files
Converted PDFs and office documents
Support articles and FAQs

Text is divided into overlapping, sentence-aware chunks before being embedded. Chunk size and overlap can be configured when the Knowledgebase is created, allowing retrieval behavior to match the source material.

Images

Multimodal Knowledgebases use Qwen3-VL to understand both text and visual content.

Images are represented by their meaning rather than their filename or manually assigned tags. This makes searches such as these possible:

"Find the architecture diagram showing the authentication flow."
"Show me the screenshot with the failed deployment."
"Which product image contains a red control panel?"
"Find the scanned form mentioning a workplace injury."

Image results retain their media type and source URL, allowing applications to display the original image alongside the answer.

Video

The same multimodal model supports video libraries.

A video can be indexed as one item or divided into time-based segments. Segmenting long recordings makes it possible to retrieve the specific clip relevant to a question rather than returning an entire recording.

Example searches include:

"Where does the presenter explain the deployment process?"
"Find the safety demonstration involving the pressure valve."
"Which training clip covers customer escalation?"
"Show the section where the speaker discusses pricing."

Video ingestion is asynchronous and durable. GraphN stores searchable vectors in Qdrant while keeping generated clips in the storage service and refreshing access URLs when results are returned.

How ingestion works

GraphN offers several ingestion paths for different workloads.

Direct upload

Text content can be uploaded directly. The service streams the document through its chunking and embedding pipeline, avoiding the need to load an entire large file into memory.

URL ingestion

Images, videos, and remotely stored documents can be ingested using a URL. A common workflow is:

Upload a file to GraphN Storage.
Generate a secure URL.
Submit that URL to the Knowledgebase.
Let GraphN detect and process the content type.

Remote fetching includes host restrictions, DNS validation, redirect checks, concurrency controls, and size limits to protect the service from unsafe or unexpectedly large downloads.

Asynchronous batch ingestion

For larger collections, a single job can ingest up to 100 items.

Each item is processed independently through a durable workflow. Clients can inspect job and item status, filter failures, retry safely with idempotency keys, or cancel outstanding work.

This makes batch ingestion appropriate for:

Importing an existing document library
Processing folders uploaded by a user
Synchronizing content from another system
Indexing many images or videos
Running ingestion without holding an HTTP connection open

Retrieval is more than vector similarity

A search begins by embedding the user's query with the same model used by the Knowledgebase. Qdrant then finds semantically similar chunks or media items.

GraphN can apply a second-stage reranker before returning results:

Text Knowledgebases use a text cross-encoder.
Multimodal Knowledgebases use a vision-language reranker.

The initial vector search is fast and broad. Reranking gives the most promising candidates a more precise relevance score.

Applications can also control:

The number of results
Minimum relevance thresholds
Whether reranking is enabled
Metadata filters
The embedding model used by the Knowledgebase

If reranking is temporarily unavailable, search safely falls back to vector order instead of failing the entire request.

Knowledgebases are designed for workflows

Retrieval becomes especially useful when combined with GraphN agents, functions, and workflow steps.

1. Company knowledge search

Ingest company documentation, handbooks, technical guides, and converted PDFs. When a user asks a question, the workflow searches the Knowledgebase and gives the relevant passages to a synthesis agent.

The result is an internal assistant that can answer questions with citations instead of relying on the model's general training data.

"What is our process for requesting production database access?"

2. Deep research assistant

A research agent can perform multiple searches, compare evidence, identify conflicting claims, and produce a structured response.

This works well for:

Literature reviews
Market research
Competitive analysis
Regulatory research
Technical investigations

The Knowledgebase supplies evidence; the agent decides how to explore and synthesize it.

3. Structured document extraction

Retrieval does not need to end with a conversational answer. A workflow can search a document collection and extract information into a predictable schema or table.

Scroll horizontally to compare

Document	Owner	Deadline	Risk	Required action

This pattern is useful for contracts, claims, medical-legal reports, product briefs, and compliance documents.

4. Contract and proposal generation

A workflow can ingest reference contracts, requirements, and supporting evidence before generating individual sections of a proposal.

Each section can retrieve its own evidence, use a specialized drafting agent, and then pass the result to an assembly step.

This keeps generated content grounded in the source materials while allowing large documents to be built in parallel.

5. Visual knowledge search

A multimodal Knowledgebase can act as a semantic catalog for diagrams, screenshots, scanned forms, product images, or field photographs.

A workflow can retrieve the most relevant images and send them to a vision agent for deeper analysis:

Search for diagrams related to network isolation.
Analyze the matching images.
Compare them with a written architecture specification.
Produce a summary of discrepancies.

6. Searchable video libraries

Training sessions, lectures, demonstrations, and recorded meetings can be divided into searchable clips.

A workflow can retrieve the relevant segments, summarize them, and return links to the exact portions of the source video. This turns large media archives into interactive learning or support systems.

7. Knowledge-grounded voice agents

GraphN voice agents can attach a Knowledgebase to live conversations.

Before generating a response, the voice system searches for relevant policies, product details, or troubleshooting instructions and adds them to the agent's context.

This enables grounded voice experiences for:

Customer support
Employee help desks
Product guidance
Training assistants
Policy and benefits questions

Built for production workloads

The Knowledgebase service separates synchronous API traffic from asynchronous ingestion workers so each can scale independently.

In production, the platform combines:

DocumentDB for Knowledgebase, document, and job metadata
High-availability Qdrant for vector storage and retrieval
Durable asynchronous ingestion
GraphN Storage for persistent media and video segments
Dedicated embedding and reranking models
Autoscaled API and worker deployments

Every Knowledgebase remains scoped to its workspace. Authentication supports both user tokens and API keys, while write operations pass through credit admission controls.

Content can also be evaluated by GraphN Guardrails during ingestion. Text is sampled across the document, while images and videos can be evaluated through their media URLs.

Complete lifecycle management

Knowledgebases include APIs and UI support for:

Creating and listing Knowledgebases
Discovering available embedding models
Uploading and listing documents
Batch ingestion
Tracking ingestion jobs and individual items
Semantic search and reranking
Metadata filtering
Updating document metadata
Batch get, update, and delete operations
Deleting documents and their vectors
Deleting an entire Knowledgebase and associated content
Viewing document and vector statistics

The embedding model, chunk size, and overlap are fixed when a Knowledgebase is created. This ensures every vector in the collection remains compatible and retrieval behavior stays consistent over time.

A reusable memory layer for AI systems

A Knowledgebase is more than a document search endpoint. It is a reusable memory layer that can be shared by agents, workflows, voice experiences, APIs, and user interfaces.

The same indexed content can support:

Conversational answers
Research and synthesis
Structured extraction
Document generation
Visual discovery
Video retrieval
Automated decision support
Real-time voice conversations

By combining multimodal embeddings, durable ingestion, semantic retrieval, reranking, and workflow orchestration, GraphN makes private knowledge usable wherever an AI system needs accurate context.

The model supplies the reasoning. The Knowledgebase supplies the evidence.