Back to articles

Tag archive

#fieldnotes

Retrieval Depth Is a Dial, Not a Doctrine
Nov 16, 2026

Retrieval Depth Is a Dial, Not a Doctrine

AWS Bedrock now lets one parameter choose how hard retrieval works on each request, from passages only to several rounds of planned searching. The setting matters less than the discipline behind it: price every query class, default low, and make the expensive mode earn its place.

Nov 16, 20266 min read0 reactions0 comments
The Second Deployment Is the Product
Nov 15, 2026

The Second Deployment Is the Product

A first deployment teaches an FDE team what a client's work actually looks like. The second is where that learning either becomes an asset the next team can pick up — eval sets, failure catalogues, connector patterns — or evaporates with the people who learned it.

Nov 15, 20265 min read0 reactions0 comments
The Reranking Bake-Off
Nov 14, 2026

The Reranking Bake-Off

We pointed two retrieval setups at our own published writing: 230 chunks, 24 questions with known answers, vector search alone against vector search plus a cross-encoder reranker. The second stage rescued buried answers, charged heavily for them, and lost on the questions where neighbouring posts disagree.

Nov 14, 20267 min read0 reactions0 comments
Compile the Knowledge at Index Time
Nov 12, 2026

Compile the Knowledge at Index Time

Chunk-and-embed makes every question redo the understanding work. RAGFlow 1.0 compiles documents into propositions with verbatim quotes at ingestion instead — and its own results say the tree was never the point; the evidence on its leaves was.

Nov 12, 20265 min read0 reactions0 comments
Circuit Breakers for Agent Tools
Nov 11, 2026

Circuit Breakers for Agent Tools

When a tool an agent depends on starts failing, the agent will keep calling it, replanning, rewording and retrying long after a person would have stopped. A circuit breaker is the part that remembers the failures and says no on the tool's behalf.

Nov 11, 20266 min read1 reactions0 comments
Agent Runs Need Deadlines
Nov 9, 2026

Agent Runs Need Deadlines

Agent runs usually carry limits on steps and spend, but no point in time after which the answer stops being useful. A deadline set from the task itself, and carried through every step and tool call, changes what a run does with its last minutes.

Nov 9, 20266 min read0 reactions0 comments
Compaction Is a Lossy Handoff
Nov 10, 2026

Compaction Is a Lossy Handoff

When an agent's history is summarized so the run can continue, something is always left out. Treating that summary as a handoff — a keep list written in advance, pointers back to the original, and a check on what survived — is what makes the trade safe to take.

Nov 10, 20266 min read0 reactions0 comments
Async Agents Move the Failure Modes
Nov 8, 2026

Async Agents Move the Failure Modes

Splitting an agent into an orchestrator and background workers keeps the conversation responsive. It also means work now fails, waits, and finishes while nobody is watching — and supervision has to be designed for exactly that.

Nov 8, 20265 min read0 reactions0 comments
Cancelling an Agent Run Cleanly
Nov 7, 2026

Cancelling an Agent Run Cleanly

Pressing stop on an agent run only clears the screen unless cancellation was designed in. The signal has to reach every tool call and child run, the call in flight has to say what it did, and cleanup has to run after the run itself is already cancelled.

Nov 7, 20265 min read0 reactions0 comments