Knowledge base

The knowledge base is what your agent answers from. It's the difference between a chatbot that guesses and a support agent that gives correct, traceable answers backed by your own material. Everything in this page is about one goal: keeping replies grounded in your content and cited back to their source.

The knowledge layer is built on FastGPT. Material is organized into datasets — collections you import, chunk, and index — and an agent retrieves from the datasets you attach to it.

Importing content

You add material to a dataset by importing it. cove1 supports several sources:

• Documents — help-center articles, product manuals, policies, FAQs, and internal guides. Paste or upload the text and it becomes part of the dataset.
• Links — point at a web page or documentation URL and cove1 fetches and imports its content, so published pages you already maintain can feed the agent.
• Files — upload files such as PDFs, Word documents, Markdown, or spreadsheets. cove1 extracts the text and imports it.

When you import a document, cove1 splits it into chunks — smaller passages sized for retrieval — and builds an index over them. Chunking matters: a well-split document retrieves precise, relevant passages, while a single wall of text retrieves noisy ones. Prefer clear headings and self-contained sections so each chunk makes sense on its own.

How retrieval works

When a customer asks a question, the agent doesn't hand your entire knowledge base to the model. Instead, the retrieve step runs a search over the dataset and returns only the chunks most likely to contain the answer.

The flow is:

1. The incoming question is turned into a query.
2. cove1 searches the dataset for the chunks that best match the query by meaning, not just by keyword — so "how do I get my money back" can match a passage titled "Refund policy."
3. The top matching chunks are collected as context.
4. That context, plus the question and conversation history, is sent to the model, which is instructed to answer from the retrieved material.

This pattern — retrieve first, then reason over what was retrieved — is what keeps the agent anchored to your content instead of relying on the model's general training. If retrieval returns nothing relevant, the agent should decline or hand off rather than fill the gap with a guess.

Inspecting which chunks were used

Because retrieval is explicit, you can see exactly what an answer was built from. For a given question, cove1 shows the chunks that retrieval returned — the passages, the documents they came from, and how well each matched. This is your primary tool for debugging answer quality.

Use it to diagnose two common failure modes:

• The right chunk wasn't retrieved. The answer exists in your material but the search missed it. Usually a chunking or wording problem — the source passage may need clearer headings, or the topic may need to be stated in the customer's language. Re-split or reword the source and re-import.
• The wrong chunk was retrieved. A stale or duplicate document is outranking the correct one. Remove or update the outdated material so the current version wins.

Inspecting retrieved chunks turns "the agent gave a bad answer" into a specific, fixable observation about your knowledge base.

Keeping answers grounded and cited

Grounding is a property you maintain, not a switch you flip. A few practices keep it strong:

• Answer only from the knowledge base. Configure the agent to rely on retrieved content and to say it doesn't know — or hand off to a person — when nothing relevant comes back. A confident wrong answer is worse than an honest "let me get someone."
• Cite the source. Because each chunk carries the document it came from, replies can point back to the exact material behind them. Citations let customers verify an answer and let your team audit it.
• Keep the knowledge base current. Retrieval is only as good as what it searches. Remove outdated policies, fold duplicates together, and re-import pages when they change. Stale content is the most common cause of confidently wrong answers.
• Scope datasets to the agent. Attach only the datasets an agent should answer from. A billing agent grounded in billing material retrieves cleaner context than one pointed at everything.

Curating the knowledge base is the highest-leverage work in running a support agent: improvements here show up directly in every answer. For how retrieval fits into the overall agent flow, see Core concepts. For the structural rules that decide whether a chunk retrieves correctly — and a 12-point audit you can run against your own corpus — see Knowledge base software for AI support.