/ BlogTechnical write-ups on Model Context Protocol, local embeddings, vector databases, and client-side memory security.
Self-consistency samples k reasoning paths and takes the majority answer. It works because errors scatter and correct answers converge, and it costs k times as much.
Read full entryLicense terms, context length, and tool-calling reliability matter more than leaderboard rank. A filter for deciding which open weights deserve a GPU-hour.
Read full entryAn MCP client is a non-browser, long-lived consumer. That breaks the assumptions behind short-lived OAuth tokens, and the fix is scoped keys with rotation.
Read full entryTwo true statements about the same entity at different times are not a conflict, they are a version history. Resolution strategies and when each one is wrong.
Read full entryTechnical merit means a falsifiable risk with a named mitigation. Vague market claims get scored down. What a Phase I reviewer is scanning for on page one.
Read full entryChange the embedding model and every vector written before the change becomes noise. There is no error, no exception, just a slow decline in recall.
Read full entryFine-tuning buys token savings and format compliance. It costs you a retraining cycle every time the spec changes, which for most teams is weekly.
Read full entryClient-side encryption kills server-side semantic search. Encrypted vector indexes, PIR, and the honest accounting of what each one costs in latency.
Read full entryReasoning models spend tokens before answering, which trades latency and cost for accuracy on search-shaped problems. A decision rule based on verifiability.
Read full entryInference is stateless by design. Every "conversation" is the full transcript replayed, and the tab boundary is just where the client stops replaying it.
Read full entryA system prompt is code with no type checker and no test suite unless you build one. Version it, pin it per release, and diff behavior before you ship a wording change.
Read full entryWeights are not uniformly important. Per-group scales, outlier channels held in higher precision, and why 4-bit lands closer to FP16 than the bit count suggests.
Read full entryRe-pasting project context into a fresh session is a measurable tax: tokens, minutes, and the details you forget to include. Here is how to actually price it.
Read full entryFunction calling is constrained decoding against a JSON Schema. Everything agentic is a loop around that primitive, and the loop is where reliability goes to die.
Read full entryA 4-bit 8B model needs roughly 5 GB of RAM and runs at readable speed on a laptop. Memory bandwidth, not FLOPs, is the constraint that decides what ships locally.
Read full entryRecall@10 hides the failure your users feel. Measure answer correctness under a fixed token budget, plus contradiction rate over multi-session traces.
Read full entryA log-sized inclusion proof lets a user verify their memory was never silently edited, without the provider handing over the whole database. Borrowed straight from RFC 6962.
Read full entryMulti-hop questions are where embeddings fall apart, because no single chunk contains the answer. The queries that justify an edge table, and the ones that do not.
Read full entryChaining multiplies per-step reliability: five 95% steps give you 77% end to end. When that math argues for one big prompt instead, and when it argues against.
Read full entryPostgres row-level security beats a WHERE clause you have to remember to write. What that costs on a pgvector query plan, and how to keep tenant leakage impossible by construction.
Read full entryAES-256-GCM with a per-ciphertext scrypt-derived key, and the awkward part nobody mentions: encrypted rows break the vector index you were counting on.
Read full entryAn asynchronous pass that dedupes, merges, and demotes memories keeps a store usable past 10,000 entries. The write policy, the merge rules, and the failure modes.
Read full entryA context window is a buffer you refill every request. Memory is state that survives the request. Conflating them is why long-context models still feel amnesiac.
Read full entryBM25 catches the identifiers embeddings smear. Reciprocal Rank Fusion merges the two without tuning weights. Where a graph hop earns its cost.
Read full entryAn exponential half-life applied to retrieval scores fixes more relevance bugs than a better embedding model. How to tune the decay constant without erasing durable facts.
Read full entryMCP is a JSON-RPC 2.0 transport with a capability handshake. What that buys you, what it does not, and why the N×M integration problem was the real target.
Read full entryCosine similarity retrieves what looks like your query, not what you need to know. A breakdown of where pure vector recall fails and what has to sit on top of it.
Read full entryBrowse our catalog of 50+ tested AI prompt profiles covering DevOps, data modeling, agent behaviors, and API wrappers. Completely free, no registration required. Browse prompt libraries →