| name | knowledge-base |
| description | Use when crawling docs, ingesting source families, building or refreshing a Qdrant-backed RAG corpus, capturing docs sections, or producing LLM-ready markdown. |
Axon Knowledge Base
Use this to turn URLs or topics into organized LLM-ready content.
Onboarding Interview
Infer the source, goal, depth, and output location from context. If the source and goal are clear, proceed immediately.
Ask at most 1-3 concise questions only if blocked, such as the source URL/topic, whether the output is reference/RAG/training/docs, or training format if training is requested.
Axon Collection Plan
Choose the Axon surface by source shape:
All indexing goes through the one unified source command — axon <source>
(equivalently axon source <source>), narrowed with --scope when the family
default is wrong. There is no axon crawl, axon ingest, axon embed, or
axon refresh.
| Need | Axon surface |
|---|
| Discover URLs only | axon map <url> |
| Capture a docs site or section as artifacts | axon <url> --scope site --output-dir <dir> |
| Fetch selected pages | axon scrape <url> --output-dir <dir> |
| Discover topic sources | axon search or axon research |
| Index repos, feeds, Reddit, YouTube, packages, local paths | axon <source> |
| Index AI sessions | axon sessions, or axon "session:<provider>:<path>" |
| Refresh an existing source | re-run axon <source> — the ledger diffs manifests and only re-prepares what changed |
| Schedule freshness | axon watch create "<source>" --every-seconds <N> |
| Reuse indexed content | axon query, axon ask, axon retrieve, axon sources |
Choose an explicit workflow output directory and pass it with --output-dir or
--output for commands that write files. Do not treat repo-local .axon/ paths
as Axon's internal data directory.
For finished knowledge-base deliverables, run indexing in blocking mode:
axon <source> ... --wait true. Without --wait true the CLI enqueues and
returns a job id — if you intentionally enqueue, report the job id and the exact
follow-up command (axon jobs get <job-id>).
Parallel Work
If appropriate, use sub-agents or equivalent parallel task runners:
- one docs section per researcher
- official docs, tutorials, community discussions, and references by source type
- source scraping vs chunk generation vs manifest generation
Output Modes
- Reference: generated markdown files,
index.md, and sources.json.
- RAG: scraped markdown/HTML/JSON files, embedded Qdrant collection,
sources,
retrieve, and optional generated manifest files.
- Training: scraped source files plus optional agent-generated JSONL/metadata.
- Docs corpus: curated markdown artifacts, source index, and table of contents.
Final Deliverable
# Knowledge Base: [Source]
## Summary
[What was collected and why]
## Output Structure
[Files/directories created]
## Coverage
[Sections, source types, counts]
## Usage Notes
[How to use in RAG, docs, training, or agent context]
## Sources
[URLs collected]
## Rerun Inputs
workflow: knowledge-base
source: [url/topic]
goal: [reference/rag/train/docs]
depth: [quick/thorough/exhaustive]
output_dir: [explicit path]
Quality Bar
- Preserve code examples and formatting.
- Remove boilerplate navigation where possible.
- Include source URLs in frontmatter or metadata.
