LightRAG setup for knowledge search in small business
How LightRAG builds a knowledge graph from business documents, and how MILabs sets it up with the owner.
Updated 2026-02-06
What LightRAG is
LightRAG is an open-source retrieval framework published by HKUDS. It reads a collection of documents, picks out the entities and the relationships between them, and stores the result as a knowledge graph. When a question arrives, the system searches that graph alongside the original text and passes the material it finds to a language model, which writes the answer. The project is written in Python, and its source and documentation sit on GitHub, where the maintainers describe how to run it on a local machine and as a server with a web interface. Because it is open source, a business can host it on hardware or in a cloud account it controls and decide which language model receives the retrieved passages. The software does not judge whether the documents are correct, so the quality of the source material shapes every answer as much as the tool does.
LightRAG in plain language
Most search tools match words. A query for "termination notice" finds documents containing those exact words and misses a clause that says an agreement may end after thirty days written warning. Questions that depend on how a client, a supplier and a project connect across several files often receive only part of the picture.
A knowledge graph holds things as nodes, such as people, companies, products or policies, and records the links between them as edges, such as "supplies", "approved by" or "replaces". Once those links exist, a question can follow a chain from one document to the next. Asking which suppliers were involved in a project later reviewed by a particular engineer becomes a walk through connected records instead of a search across unrelated pages.
Building that structure is done by a model, which reads each passage and suggests the nodes and links. The result is a graph a team can review and refine, and the source passages stay visible beside each answer, so every connection can be traced back to its evidence.
That makes the tool suited to questions about connections between documents. It serves least well where a team only needs one file found by its title.
What businesses use it for
- Searching a library of internal policies, procedures and templates by meaning rather than exact keywords
- Answering staff questions about contracts, pricing rules or compliance steps from approved documents
- Building a map of how clients, suppliers and projects relate across long document archives
- Preparing briefings from a set of reports, minutes or research papers the business already holds
- Keeping a private question-answering service running on company hardware for sensitive files
- Checking draft client communications against the policies they must follow
Who it suits
- Firms with large document archives that staff search every week
- Businesses that must keep client files on hardware they control
- Teams comfortable with a self-hosted tool and a named person to maintain it
How MILabs sets it up
MILabs begins with the owner or team lead, agreeing which documents the knowledge base should hold and which of those are current. The team then decides where LightRAG will run, either on hardware the business already controls or in a cloud account the business owns, and selects the language model that will receive retrieved passages. Next, the document store is connected to the folders, shared drives or client files the business already uses, so new material can be indexed without manual copying. Access is then set so staff see only the material their role requires, and the business confirms who reviews answers and who updates documents when policies change. MILabs is a Sydney agency and delivers this work remotely to businesses in other cities.
Builds it fits
Industries it fits
LightRAG in practice
In a typical engagement the work begins with a sorting exercise. Staff gather the documents that answer recurring questions, separate the current versions from the superseded ones, and mark which files carry authority, such as a signed policy against a draft circulated by email. That sorting sets the boundary of the tool, so it deserves as much attention as the software itself.
Once the collection is settled, the team runs a set of test questions drawn from real work. A compliance officer might ask which steps apply before a contract variation, or a senior accountant might ask which client agreements mention a particular fee structure. Each answer is read against its cited passages, and the team adjusts the extraction or the source files whenever the graph joins documents that should stay apart or misses a link a reviewer expects.
Access comes next, with the business deciding which roles may query which folders. Questions staff cannot yet answer get logged as they arise, and that running record becomes the agenda for each review. Over time the log shows whether the gaps lie in the documents, in the questions people ask, or in the way answers are presented.
Frequently asked questions
- Can LightRAG connect to the software I already use?
- Yes, the usual route is to point it at the folders, exports or shared drives where the documents already live. Whether it reads a particular platform directly depends on the connector or script used to export the content, and MILabs checks that route with the owner at the outset. Records that change often need a scheduled update so the knowledge graph keeps pace with them.
- How does LightRAG compare with other tools for the same job?
- A plain document chat tool such as AnythingLLM suits questions answered by a single file. LightRAG is more useful when answers depend on how many documents relate to each other. The best test is to run both tools against the same set of questions using the business's own documents, then compare which answers are accurate and which sources they rely on.
- How do I get my team to use LightRAG?
- Start with a small group who already search the same documents each week, and show them the answers alongside the source passages. Invite them to log wrong or missing answers in a shared list, and name one person to keep the document set current. Adoption usually follows when the first few answers save visible effort in the working week.
- Can I self-host LightRAG?
- Yes. LightRAG is open source, and its project documentation describes running it on your own machine or server. Self-hosting keeps the documents and the knowledge graph on hardware the business controls. The language model it calls is a separate choice, and a local model keeps prompts on-site too, though it may answer less well than a hosted one.
- Which builds work well with LightRAG?
- Builds work well when staff need answers grounded in a fixed set of documents, such as the staff knowledge assistant, the document question-and-answer assistant, the private document server and the policy compliance assistant. A client FAQ assistant also suits it when the answers come from approved material. The right choice depends on who will ask the questions and what they are allowed to see.
