What you are building
An authoritative enterprise knowledge assistant that answers technical, operational, and organizational questions for employees across every department:
- "What is the step-by-step procedure for rolling back a canary deployment in production?"
- "What is our updated expense reimbursement limit for customer dinners?"
- "How do we configure single sign-on (SSO) for new contractors joining the data team?"
- "Who approved the updated security incident response guidelines?"
In growing organizations, documentation is spread across hundreds of separate repositories, wikis, and cloud folders. / ml synthesizes this scattered corpus into a coherent knowledge graph that reflects your current operating reality.
Why memory decides whether it works
Standard vector search breaks down on internal docs because wikis constantly change:
- Old docs outrank new updates: A long 3-year-old runbook has more keyword matches than yesterday's short update, so RAG serves the outdated procedure.
- Guessing when docs are missing: When no policy exists, vector search returns weak matches and the model invents an answer.
- Missing citations: Engineers cannot trust an operational answer without a direct link to the source doc and revision.
- Draft bleed: Unfinished notes mix with approved company policies unless scoped by status and department.
What / ml provides
Easy Doc Updates & Cleanup
Update changed procedures in place, merge duplicate guides, or delete retired docs in 6.41 ms so old instructions never resurface.
No Guessing When Docs Are Missing
Set a minimum match score so weak search results are dropped, letting your assistant say "I don't know" when a policy doesn't exist.
Direct Link to Every Doc
Every answer includes the page URL, section title, and update history so employees can open the source doc in one click.
Department & Role Filters
Tag docs by team, project, or access level so each department only searches policies they are allowed to read.
Cost and token economics
Standard search pastes 5 to 10 long document chunks into every prompt, burning over 7,000 tokens even for simple policy questions.
Tokens Used Per Search
Pasting long doc chunks: ~7,400 tokens ($0.037 per question)
/ ml relevant facts only: ~1,250 tokens ($0.0062 per question)
Result: 83% lower token cost and cleaner, faster answers.
For a 1,000-person company asking 150,000 internal questions a month, / ml saves over $4,600 monthly in model costs.
Getting there
Sync your docs and runbooks
Send handbooks, wiki pages, and markdown files from GitHub, Notion, or Confluence to / ml with their page URL and last-updated date.
Tag docs by team and access level
Add simple tags like department: engineering or status: approved so drafts and private docs stay separated.
Search with team filters and match scores
When an employee asks a question, search / ml with their team filter and a minimum score so weak matches are ignored.
Include the doc link in every answer
Have your assistant answer strictly from the returned facts and show the doc_url link so readers can check the source page.
Code example
Ingesting updated infrastructure runbooks and querying with metadata filters and source attribution:
import os
from slashml import Memory
mem = Memory(api_key=os.environ["SLASHML_API_KEY"])
# 1. Ingest an updated production rollback runbook
mem.add(
"Production Rollback Runbook: To revert a canary deploy, run `kubectl rollout undo deployment/api-server -n prod`. Do not restart pods manually.",
user_id="company-kb",
metadata={
"doc_url": "https://wiki.internal.net/runbooks/rollback-v3",
"title": "Production Rollback v3",
"department": "engineering",
"updated_at": "2026-10-25"
}
)
# 2. Query company brain for active procedure with metadata filter
result = mem.search(
"How do I rollback a canary deploy in production?",
user_id="company-kb",
filters={"department": "engineering"}
)
if not result["results"]:
print("I cannot find documented instructions in our company knowledge base.")
else:
for item in result["results"]:
meta = item.get("metadata") or {}
print(f"Procedure: {item['memory']} (score: {item['score']:.3f})")
print(f"Last Updated: {meta.get('updated_at')}")
print(f"Documentation Link: {meta.get('doc_url')}")Frequently asked questions
Q How do we replace an outdated runbook or policy?
You can update a doc in place (which saves the previous version in its history), merge overlapping pages, or delete retired docs in 6.41 ms.
Q How do we separate docs across different departments?
Tag each doc with its team, project, or access level. When someone searches, filter by their department so they only see docs meant for them.
Q How do we connect Confluence, Notion, or GitHub?
Your sync script or CI action sends markdown or plain text directly to / ml using our Python SDK, TypeScript SDK, or REST API, along with the page URL and date.
Q How does / ml stop the assistant from guessing when no doc exists?
Standard vector search always returns the closest text even when it's unrelated. / ml lets you set a minimum match score and rerank results, so weak matches are dropped before they ever reach your model.