Verify intended redactions were actually redacted before publishing documents
The speaker recommends verifying that whatever was intended to be redacted in a document was actually redacted before publishing it, to avoid legal exposure.
35,684 results · Most relevant
The speaker recommends verifying that whatever was intended to be redacted in a document was actually redacted before publishing it, to avoid legal exposure.
Run the improvement loop (identify issues, build evals, make changes, run experiments, ship) continuously rather than once. Agent performance degrades over time due to model regressions and changing user question patterns — even two months after a working fix — so set up evals to monitor and observe degradation, and build evaluations to continuously monitor use cases after fixing identified issues.
Nyah Macklin walks through creating a Neo4j Aura instance for the workshop, saving its connection details, configuring the .env file with the Neo4j and OpenAI credentials, and running testenvironment.py to verify the setup.
To follow along with the workshop, attendees create a Graph Academy account (graphacademy.neo4j.com), a completely free forever-free online learning platform covering concepts like building agents, agentic memory, and context graphs. They then create a free Neo4j Aura instance (not a paid one), which allows hooking up to Aura agents and building agents within the Aura platform; if you want more tools in the future, you can feel free to click the professional (paid) Aura instance instead. The workshop also provides a free OpenAI API key that gives access to build during the session — copy it down or save it on your computer for later access. The key has usage limits, so you cannot run up the bill. All links and notes are at dev.neo4j.com/workshop-services gen-ai dash-notes, with slides at 9
If you can traverse your knowledge graph and use that plus things like agent memory, you get rich historical context. Traceable and explainable agent decisions are especially important in regulated domains like finance, medicine, and legal. You need to move fast, be accurate, and be explainable when building applications, and AI governance and safety matter when thinking about how your agent makes decisions and how it is built.
Bostrom recommends prioritizing defensive technologies to arrive before dangerous ones — for example, in biotechnology, getting vaccines, prophylactics, and diagnostics before bioweapons, even when the same underlying technology unlocks both.
attributed
As a leader, create an environment where people can ask uncomfortable questions and be able to say 'hey, I don't think this is working.' Create a culture where the team can call out both good things and bad things. Trust the team a lot because they're the ones every day trying to make the thing go. If the team says 'hey, it's not going,' then the response is: okay, let's pivot, let's do something different.
Josh Woodward describes a 'big thank you' tradition: leaders put a surprise meeting with no agenda on someone's calendar, start by saying 'this is not a bad meeting. This is a good meeting,' and then go around the circle telling the person why they appreciate them. He notes this works because people usually assume the worst when they see a scary meeting invite, so the ritual deliberately does the opposite.
Remy recommends being able to make a choice at any given layer that you need at any given time, noting that the choice of AI components might be different for different teams or folks across the organization.
For the query plan step, pass the schema to the LLM, give it a tool, and force it to call that tool for query synthesis, then execute the plan. The query plan step is arguably the simplest part because it's the LLM part. LLMs are actually pretty good at generating temporal queries. Pass the current date and time to the LLM in the prompt so it can generate temporal queries accurately.
Separate MongoDB collections by tenant instead of housing multiple tenants in one collection. Do not house multiple tenants in a single Mongo collection.
Create a full-text search index and combine it with author and state filters in the same match clause. When regular expressions and FTS are disallowed, the LLM falls back to exact match, which may fail to find results like 'authentication fails' — add full text search to handle fuzzy queries.
Build the system starting with a full UX application and a stubbed engine, then gradually build it up; at step three the thing comes alive and you can issue queries to it. Any repo is fine for the workshop setup. As a shortcut, you can point an agent at the repo and let it set everything up for you. The workshop follows a series of six pull requests that stack on top of each other. If anyone needs help during setup, put your hand up and the presenter will walk over to your desk to help. If you don't get all the steps nailed down as you go, don't worry, because the repo is online and you can run through it at your leisure afterwards.
Identity resolution can be done dynamically: have the LLM synthesize the identity mappings, though coupling to the schema at first blush seems likely. For identity resolution in this project, use simple fuzzy matching such as Levenshtein distance across names.
partial
To scale across multiple collections, add another LLM step on top where schema collections have tiny summaries and descriptions, then load the right schema depending on the query. LLMs are surprisingly good at selecting the right schema for a query now, though they weren't a year ago.
Before starting a running training block, check whether the person has trained enough to actually train running; 'training to train' is a distinct phase from a running block.
To connect a client: add a server, choose 'remote', and replace the URL with the URL received from deployment — this works with any client that supports MCP. After connecting, the five CRUD tools defined by the MCP are visible via 'view tools', and you can chat with your IDE in the context of your project so it can leverage the MCP server. In the demo, the IDE identified the intent to track a decision to use OAuth 2.0, asked permission to run the insert decision tool, and the call to Akamai Functions tracked the decision.
The speaker advises that if cost is a concern, build reusable agents and have people solve problems the same way rather than reinventing the wheel, to drive down your overall bill. Reusable components drive up cache rate because you get exact matches between turns on common data.
Mansour advises that when building features for humans, teams should also expose them as skills and tools for agents, based on Atlassian's experience converting human features into agent-usable tools.
Pedro Lopez advises that when implementing OAuth for MCP, keep offline long sessions stored separately from the web app with their own set of life cycles, since users expect long-lived sessions through MCP servers — more like a mobile app where they don't have to think about authenticating.
Choose sources → to get a library of your own and a daily brief.