Knowledge
Follow a claim back to the evidence.
Watch the 70-second film: a claim that lost its source, and how the tool puts the quote and the date back on it. The example in it is fictional.
For people connecting notes, documents and research. Knowledge turns supplied material into an inspectable graph, with sources and time attached, so you can trace relationships, question an apparent contradiction and see what changed.
Explore the site
What you get
- Keep the source within reach. Carry provenance with the claim instead of relying on an answer with no path back to its inputs.
- Explore a specific question. Use paths, evidence, gaps and timelines to navigate relationships rather than rereading an entire collection.
- Keep a history of changing claims. Close a relation’s validity window and append its replacement instead of erasing the earlier view.
Start here
git clone https://github.com/michaeljabbour/amplifier-capability-knowledge.git
cd amplifier-capability-knowledge
python3 scripts/capability.py install --dest ~/.agents/skills
This installs the portable method and its packaged references into your agent skill folder. If your host uses another folder, pass that path to --dest. Start a new session or use your host’s skill refresh so it can discover the installed skill. Optional native tools are installed separately.
Try this first task:
Use Knowledge to show how two fictional project notes support or challenge the same claim. Start with deterministic ingestion or structured additions. Trace each relationship to its source and distinguish an asserted fact from a candidate contradiction. Do not configure an external provider.
What you need
- Python 3.11 or later for the portable install helper.
- The deterministic library and CLI use Python and SQLite without a model key.
- Native package and command identity stay knowledge; optional model-backed operations need a separately configured route.
- Choose KNOWLEDGE_HOME explicitly when you want a disposable test graph.
For the optional native interface, use a project virtual environment:
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install .
Evidence and limits
Native graph, storage, analysis and interface tests passed. A passing graph operation does not establish that an extracted claim or model-written answer is correct.
The worked examples on the site are fictional. They illustrate the output and decisions; they are not user results or measured benefit. See consolidation and runtime boundaries for retained interfaces and acceptance scope.
Retained source guide
knowledge
Ask your documents a question. See the answer as a map.
Point it at papers, specs, policies, notes, code or agent sessions. You get a graph any agent can
query and add to: concepts and typed relations, each with the passage it came from, the agent that
added it, a confidence and the dates it held. Ask in plain words and it answers in the form that fits:
a timeline of what changed and what replaced it, an argument with the quotes for and against a
claim, the steps between two ideas, a matrix of what is missing, a rhythm of when each topic was
written about, or the map. When a question is unclear, it asks you instead of guessing.
It is an Amplifier Smart Tool: a Python
library with a thin CLI and an optional MCP server, usable from Claude Code, Codex, Amplifier, a
script or any agent. Everything except five model-backed commands runs with no model and no key.
Product page and demos ·
One-minute trailer · How it works
Four Claude Code agents reading Amplifier's public docs and writing to one graph at the same time,
recorded live with knowledge serve (sped up).
Quick start (no model, no key)
uv tool install "amplifier-smart-tool-visualizer @ git+https://github.com/michaeljabbour/amplifier-capability-knowledge"
knowledge doctor
knowledge ingest ./notes --method cooccurrence --graph notes # a word network, no model
knowledge analyze --graph notes # central ideas, bridges, topics, gaps
knowledge visualize --for "what are the main topics?" --graph notes --open # opens on the answer
With a model, knowledge ingest ./notes extracts typed concepts and relations instead (Anthropic,
OpenAI, a local Ollama model, or your own program through --complete-cmd). Inside an agent, the
agent is the model: ingest --method agent hands it the chunks and the extraction prompt, and it
writes back with knowledge add. No key is billed.
Examples you can open now
Real public material, each read into a graph by an agent through the tool's own agent route, each
view opening on a question someone in that job would actually ask. Every relation links to its quote.
| For | Material | Ask it | What you see |
|---|
| Developers | Python's typing PEPs: 484, 526, 563 (four versions), 585, 604, 649, 749. Public domain / CC0. | "What replaced PEP 563's plan to make stringized annotations the default?" | A timeline: the plan moved from Python 4.0 to 3.10, was put on hold for PEP 649, and "never became the default behaviour"; hatched bars point to PEP 649 and 749's deferred evaluation. |
| Researchers | Eight arXiv abstracts on scaling laws and emergent abilities. CC0. | "What contradicts the claim that emergent abilities appear suddenly and unpredictably?" | An argument: the claim, with quote cards against (Schaeffer et al. 2023: they "evaporate with different metrics or with better statistics") and for. |
| Policy, legal and operations work | Seven US federal documents on agency use of AI, EO 14110 to OMB M-25-21 / M-25-22. Public domain. | "What did agencies have to do as of 2024-12-01?" | A timeline marked at that date: what applied then (M-24-10's December 1, 2024 deadline) and what replaced it after M-25-21 "rescinds and replaces" M-24-10. |
| |
|---|
| Developers. A timeline of a plan that was replaced. | Researchers. An argument: a claim and the quotes against and for it. |
| Policy work. What applied on a given day, and what applies now. | Six forms, one graph. Map, timeline, argument, steps, matrix, rhythm. |
| What's missing, as a matrix. Gap cells outlined, with their bridge candidates. | It asks, not guesses. A vague question gets a question back, with the command for each answer. |
Build them yourself: python3 scripts/build-examples.py (no model, no key; it checks every question
still opens the view it should).
What it found in Amplifier's own docs
The example graph is real: twenty public documents from
microsoft/amplifier-docs and the
smart-tools spec, read by four
agents into 313 concepts and 613 relations, each citing an exact quote.
- The biggest gap: the smart-tools world (cli, smart tool, manifest) has almost no links to the
kernel world (kernel, mount plan, module).
knowledge gaps scores it 0.97 of a possible 1.
- Tensions in the text: the profiles guide says inheritance never removes parent features, then
shows a profile that drops
tool-bash. The tool contract says execute() returns ToolResult;
its example returns dicts. The spec asks callers not to care whether a result came from a model,
and also asks tools to say which capabilities use one. knowledge contradictions lists 13.
- Beliefs that changed: in amplifier-foundation's history, token
streaming was enabled by default, paused as "temporary" the same day, and re-enabled the next.
knowledge timeline "token streaming" shows all three, with the dates each held.
Demos
| |
|---|
| Find the gaps. Topics by colour; the gap list names the biggest holes and the concepts most likely to bridge them. | Check a claim. Supporting and contradicting relations in green and red, each with its quote, agent and date. |
| Trace a line of reasoning. Shift-click two concepts; the path follows typed relations, each with its sentence. | Replay what changed. Press Play; replaced decisions stay in the history, dashed. |
| Made for agents. Plain commands, real answers, no key. | No model at all. This repository's docs as a word network. |
Capabilities
| Deterministic: no model, no key, same answer every time | Model-backed |
|---|
| Write | ingest --method cooccurrence|agent, add, relate, supersede | ingest (default method) |
| Read | search, show, path, timeline, evidence, context | ask |
| Analyse | analyze, communities, gaps, contradictions | report, questions, reconcile |
| View and share | visualize, serve, export, events | |
| Setup | stats, graphs, doctor, mcp, manifest | |
- Nothing is overwritten. Every relation has the dates it held.
supersede closes one in favour
of a newer one, a changed file closes its old claims, and --as-of DATE reads the graph as it was.
- Many agents, one graph. Agents write through
add and relate with their name on every
relation; events is the change log and serve shows it live.
- Take it anywhere.
visualize writes one HTML file with no outside requests. export writes
JSON, Cytoscape, GraphML (Gephi), CSV, an Obsidian vault with wikilinks and a Canvas map.
- For scripts and agents.
--json gives one envelope on stdout; exit codes 0, 1 (input), 2
(usage), 3 (set something up), 4 (model call failed). It never prompts.
- No surprise bills. Inside Claude Code, Codex or Amplifier, model-backed commands stop with
host_model and name the deterministic route instead of using an API key.
knowledge --help prints the tool's skill, written for an agent; knowledge <capability> --help
prints one capability in full.
In Amplifier
The tool stays a plain CLI any harness can call; Amplifier gets it through a behavior:
amplifier bundle add 'git+https://github.com/michaeljabbour/amplifier-capability-knowledge@main#subdirectory=behaviors/knowledge.yaml' --app
It loads the knowledge skill from skills/, which tells the agent to install the CLI and use the
deterministic routes. Optional: behaviors/knowledge-mcp.yaml serves the same capabilities as MCP
tools through Amplifier's MCP module, named mcp_knowledge_knowledge_<capability> (it needs
knowledge on PATH; the file shows a uvx alternative).
As a library
The library is the tool; the CLI and MCP server are thin wrappers. Results are plain dicts.
import knowledge as kg
kg.ingest("research", documents=kg.load_documents("./notes"), method="cooccurrence")
kg.relate("research", "experiment 14", "supports", "delegation reduces agency",
source_type="evidence", target_type="claim", confidence=0.7, agent="researcher")
for p in kg.path("research", "delegation", "human agency")["paths"]:
print(" -> ".join(p["nodes"]))
# your own model, no SDK needed
kg.ingest("research", text, complete=lambda system, prompt: my_model(system, prompt))
Where the ideas came from
The core is rahulnyk/knowledge_graph: chunking, an
LLM extraction prompt, contextual proximity and communities, rebuilt as a smart tool with no runtime
dependencies. Around it are capabilities taken from Graphiti (temporal edges), InfraNodus
(centrality and structural gaps), Microsoft GraphRAG (community reports and local-search context),
Obsidian tools (vault and Canvas export) and Gephi. docs/VISION.md maps every idea
to the file it lives in, and lists what was left out on purpose.
Development
uv run --with pytest python -m pytest -q
uvx ruff check --isolated --line-length 132 --select E,F,W src tests
uv run path/to/amplifier-smart-tools/conformance/run.py . # the spec's conformance kit
python3 scripts/build-examples.py # example graphs, no model
Read AGENTS.md before changing behaviour. contracts/ describes the CLI
envelope and the graph document; site/ builds the product page and demos.
License
MIT. Vendored and adapted third-party material, all MIT, is credited in NOTICE.
Repository migration
This repository was renamed from amplifier-smart-tool-knowledge to amplifier-capability-knowledge. Existing installed method names, bundle aliases and native interfaces stay the same. See the migration guide and repository map for source URL updates.
Optional coaching
Native operations remain direct. For an explicitly requested practice or review checkpoint, use the local coaching guide. It includes a fictional first task, a baseline, a bounded budget, stop rules and human-reviewed keep/revert. No learner trace is saved by default, and software checks do not establish human learning or outside-user usefulness.
