knowledge-wiki

Setting up the knowledge-wiki quickstart

Follow these steps in order when the user asks you to set up or run this quickstart. Confirm each step's output before moving to the next.

0. Check access

Everything runs on Managed Agents (beta) — memory stores, agents, sessions and dynamic workflows — which is enabled by default for API organizations; the SDK sends the beta headers itself. If client.beta.memory_stores.create or client.beta.agents.create returns a permission error later, the user should ask their organization's admin whether it was turned off or the key restricted. Python 3.11 or later is required.

1. Install dependencies

pip install -r requirements.txt

requirements.txt pins the anthropic Python SDK at >=2.0.0,<3. If the user already had an older version in the environment, confirm the upgrade took:

python3 -c "import anthropic; print(anthropic.__version__)"

If that prints a version below 2.0, run pip install -U "anthropic>=2.0.0,<3".

2. Authentication and environment

The notebook constructs Anthropic() with no credential arguments, so it resolves credentials through the standard chain and uses the first one it finds: ANTHROPIC_API_KEY, then ANTHROPIC_AUTH_TOKEN, then the profile ANTHROPIC_PROFILE names, then Workload Identity Federation, then the active ant auth login profile. Pick whichever the user's org uses:

# Either an API key…
export ANTHROPIC_API_KEY=...

# …or an interactive login with the ant CLI, version 2.0 or later (ant --version prints it),
# which stores a profile the SDK finds on its own.
ant auth login

Do not set ANTHROPIC_API_KEY alongside a profile — a stale exported key silently shadows the profile. A key set next to ANTHROPIC_AUTH_TOKEN wins too (the SDK logs a warning and uses the key). ant auth status shows which credential source won.

One more variable is required regardless:

export EDGAR_USER_AGENT="your-name your-email"  # SEC policy

If EDGAR_USER_AGENT is missing, the scripts send a placeholder contact, which SEC may throttle or refuse.

3. Pick a tier and set the cost expectation

Before fetching, tell the user what the core run cost. For other tiers, say no figure was recorded, and that the notebook prints an estimate before the build and stops if it is over budget.

Tier	Docs	Use when
core (default)	8 filings (annual reports, proxies, a 10-Q, the deal's 8-Ks), 2.8 MB text	the recorded run, 23 min and ≈ $16 at list price on the default models: the demo questions and the measured comparison are written for it
quickstart	8 short deal-window filings	a first look at the mechanics on short filings
mini	26: the smaller deal-window filings	more documents, less text than core
standard	37	the larger room
full	all 42 deal-window filings	every filing in the deal window; it leaves out core's FY2022 annual report

Only core was recorded (Claude Haiku 5.5 for the agents that read, Claude Sonnet 5.5 for the lead agent and the raw-documents baseline, thinking at the platform's default); the user's run will vary with model choice and API pricing at the time. For any tier the notebook prints its own list-price estimate for the chosen corpus before the build (the cell that creates the agents) and stops if it is over BUDGET_USD; the build session also carries that budget as a cap (the platform stops new work there, to within one model request per working agent). Empty data_room/docs before fetching a different tier: the build reads whatever is in the folder, and the fetch script stops if it finds another tier's files there.

4. Fetch the corpus

Run in this order (network required):

python3 fetch_data_room.py --tier=core    # or the tier chosen in step 3
python3 fetch_real_deck.py                # optional: 6 slides of a real board deck (~1 MB)
python3 make_analyst_docx.py              # optional: regenerates the Word note that ships in example_data/

All five manifests ship with the folder, so nothing else is needed before the fetch. build_manifest.py is an optional refresh from EDGAR: it rewrites the quickstart, mini, standard and full manifests (never core), so run it before fetch_data_room.py if at all, not after. Expected after this: data_room/docs/ contains one .txt file per document in the chosen tier — 8 files for core (about 2.8 MB) or quickstart, 26 for mini — and the notebook's normalize step adds the two synthetic files, plus the deck if it was fetched.

5. Open the notebook

jupyter lab distill_documents_into_knowledge_wiki.ipynb

Recommend the user reads README.md §"How it works" before running cells — the pipeline has a few non-obvious steps (the schema seeded into the store, one message stating the goal and naming the agents, the checks on what ran, the read-only analyst attach).

Gotchas to warn about

• The build is one long cell (Run the build): it prints each phase and thread as it starts; tell the user to kick it off and check back rather than watching it. If the kernel is interrupted, the cell's finally asks the agent to stop the run, waits for it to end, then saves the records and archives the session; the wiki store keeps what was written. If it prints that a session was not archived, archive it by hand: an open run keeps spending until it reaches the budget.
• The message states the goal and names the agents, and the lead agent writes the workflow; after the run the notebook asks the lead agent to print the workflow it wrote and sets that beside the platform's records (Check what ran). The reprint is the lead agent's account; if it and the records disagree, the records win.
• Model choice: by default the agents that read (extractors, resolver, analyzers, organizer, analysts) run on claude-haiku-5-5 and the lead agent and the raw-documents baseline on claude-sonnet-5-5. COOKBOOK_MODEL sets every stage; COOKBOOK_BUILD_MODEL (extractors, resolver), COOKBOOK_CONSOLIDATE_MODEL (analyzers, organizer), COOKBOOK_QUERY_MODEL (analysts), COOKBOOK_LEAD_MODEL (the lead agent) and COOKBOOK_BASELINE_MODEL (the baseline in "What it costs") override one stage each. The extractors and analyzers read the whole corpus and are most of the cost.
• Run records go under the system's temporary directory (/tmp/knowledge_wiki-<uid>/runs/, readable only by the user; or COOKBOOK_RECORDS_DIR), never into the repository: they hold the documents' contents.
• If the user hits a permission error on client.beta.memory_stores.create, see step 0.