Knowledge / RAG — AiHummer Docs

Knowledge is how AiHummer grounds answers in your documents rather than the model’s training data. You ingest content, the agent retrieves the relevant passages with the search_knowledge tool, and answers come back with citations. For harder questions, the deep_research tool runs a multi-step investigation and produces a cited report. Knowledge is administered under /v1/admin/knowledge/* and /v1/admin/knowledge/connectors/*.

Ingesting content

You can ingest documents, PDFs and URLs into the knowledge base. Ingested content is indexed so it can be retrieved at answer time and attributed back to its source.

[!TIP] Ingest the canonical version of a document once and let agents cite it, rather than pasting long passages into prompts. Cited retrieval keeps answers verifiable and your context small.

Grounded answers and the tools

Two tools expose the knowledge base to an agent:

Tool	What it does
`search_knowledge`	Retrieves relevant passages and grounds the answer with citations.
`deep_research`	Runs a multi-step research process across the knowledge base and produces a cited report.

Both arrive in the turn as tool results, never as injected instructions — the same prompt-injection discipline applied everywhere in AiHummer. Answers carry citations so a reader can trace a claim back to its source.

[!NOTE] deep_research is for genuine multi-step questions — it costs more time and tokens than a single search_knowledge call. Reach for it when one retrieval is not enough.

Knowledge connectors

Beyond manual ingest, knowledge can be pulled from external sources via connectors managed under /v1/admin/knowledge/connectors/*. Seven connectors are available; each one’s credentials live in the encrypted secrets vault and a background scheduler runs the syncs:

Connector	What it pulls
Google Drive (`gdrive`)	Google Drive files via a service account (credential JSON).
Microsoft Graph (`msgraph`)	SharePoint / OneDrive documents via a Microsoft Graph app.
Notion (`notion`)	Notion pages via an internal-integration token.
Slack (`slack`)	Slack channel message history via a Web API token (all channels or a configured list).
S3 (`s3`)	Objects from an S3-compatible bucket (AWS S3, GCS in S3 mode, MinIO, etc.), optionally under a prefix.
SQL (`db`)	Rows of a read-only SQL query against a database, ingested as documents.
Confluence (`confluence`)	Confluence Cloud pages (Basic auth with an Atlassian API token), optionally scoped to one space.

Separately, a one-off Slack export import takes a workspace-export ZIP uploaded manually — no stored secret and no schedule.

[!WARNING] The Google Drive connector is live-proven; the others are implemented and covered by tests, but verify them against your own data before relying on them in production.

Vector store and embeddings

By default, retrieval can run on an in-memory store with a hash embedder, which is enough to get started. For production-quality semantic retrieval, point AiHummer at a real vector store and embedder:

AIHUMMER_QDRANT_URL=http://localhost:6333
AIHUMMER_EMBEDDER_URL=http://localhost:8081

When these are set, ingestion and search_knowledge use the external vector store and embedder instead of the in-memory fallback.

[!NOTE] The semantic embedder is opt-in, not a default: the host-native installer provisions it only when you pass --with-embedder (Ollama with a light multilingual model; it then points AIHUMMER_EMBEDDER_URL at http://127.0.0.1:11434/api/embeddings). Without an embedder, retrieval runs on the lexical fallback.

Admin API

Resource	Purpose
`/v1/admin/knowledge`	Knowledge base management, including ingest
`/v1/admin/knowledge/connectors`	Configure KB connectors (Drive, Graph, Notion, Slack, S3, SQL, Confluence)

Where to next

See search_knowledge and deep_research alongside every other tool in Tools and the tools catalog.
Add per-conversation long-term recall with Memory (Einstein).