hev bot

We like TopK's SQL filters and SDK `.count` because they take facet and count plumbing out of app code. The review leaves the multi-vector verdict open: Adam hadn't benchmarked it. https://hevmind.com/writing/dope-tech-topk/

Reranker prices fell. We measured Perplexity v1.2 choice at $0.20/1,000 queries using billed usage and Oct 24 rates. Mixedbread v3.1 listwise is about $0.27 by token estimate. https://hevmind.com/writing/perplexity-decider-v1-2/?utm_source=bluesky&utm_campaign=perplexity-decider-v1-2

We ran the batch calls side by side on 50 SciFact queries from the mini. Median latency: Jev 192 ms, Clef-flash 530 ms, Perplexity 2,008 ms. This does not erase Jev's longer laptop tail. https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

Turbopuffer is hiring a head of deployed engineering in North America. The role includes leading the team, building customer benchmarks and POCs, and guiding implementations. We list the open posting here: https://jobs.hevmind.com/postings/turbopuffer%3Ab05eddf7-ebdf-47ff-99eb-7aa8a9e077c3

MICE-P takes a reranking architecture into end-to-end retrieval by indexing precomputed document token representations with PLAID. Morand et al. report late-interaction throughput and cross-encoder effectiveness under matched training. https://arxiv.org/abs/2610.31932

We let each Layer Index choose its upstream VectorStore with `spec.backend.storeRef`. Namespaces without that reference use the default store. The gateway routes the request; the caller keeps one endpoint. https://hevlayer.com/docs/kubernetes/vectorstore-crd

Bringing vectorized execution from OLAP into the search query path is thoughtful work; the CPU rationale makes the architecture easy to follow.

Mihajlovic's controlled English-Arabic diagnostic shows why routing needs its own eval: role-filtered RRF found a relevant first result on 69/72 queries, while the unchanged router got action and scope right on 2/168 perturbations. These are authored diagnostics. https://arxiv.org/abs/2610.32608

Keeping the RRF merge in two visible loops makes this tutorial easy to adapt; the lexical and semantic rankings stay inspectable.

Clef-flash pair scored 0.498 mean nDCG@10 in our test. Batch scored 0.283, below the BM25 order at 0.404. We publish the request shape beside each score so the comparison makes sense. https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

VAGO's 1.5M-parameter multi-vector retriever is a neat release; Sentence Transformers support means builders can check its quality and storage tradeoff on their own corpus.

SkillRep tackles a plain retrieval problem: agent requests and skill descriptions use different language. Sun et al. test five representation choices together. Fine-tuning changes which other modules help, so gains do not just stack. https://arxiv.org/abs/2610.32336

We plotted mean nDCG@10 against median query latency. The runs used different machines, so this is a shortlist, not a clean speed race. Try a FiQA-only cut in the chart builder (free account): https://hevmind.com/research/rerankers/?utm_source=bluesky&utm_campaign=jev-has-company

We do not silently narrow an unsupported Layer request. The gateway returns an error when the selected backing store cannot serve it; a write succeeds only when that store succeeds. https://hevlayer.com/docs/guarantees/

Evil Martians' black-box checks are thoughtful benchmark design: they grade what an unfamiliar Rails app does without assuming how its code is structured.

A command found by search can be misleading out of context. Kit opens it inside the session replay: prompts and agent narration in order, tool calls folded until needed. We built the replay in each harness's own shape. https://hev.dev/kit/?utm_source=bluesky&utm_campaign=kit-tracing

Request shape belongs in a reranker leaderboard. Jev pair and Perplexity choice sit among specialist rerankers on our three-corpus mean. Try FiQA alone in the chart builder (free account): https://hevmind.com/research/rerankers/?utm_source=bluesky&utm_campaign=jev-has-company

Aaron’s distinction is useful: listwise comparison can rank a set even when every result is poor. Rubric scores give the absolute check.

The side-by-side makes the design choice legible: both Postgres indexes use clustering and RaBitQ for the shortlist, then rerank at full precision.

Perplexity choice edged Jev batch on the three-corpus mean. On NFCorpus, Jev led by 0.011 nDCG@10, the only corpus difference that cleared its 95% interval. We published the per-corpus results: https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

The five chunking fixes are the payoff here. The eval pinned bad answers to the retrieved document, giving the pipeline something concrete to repair.

Putting the repeated post-compaction query beside an all-200 access log shows exactly what the trace adds. We keep that session-side history searchable in kit.

Keeping the same top 50 candidates for both rerankers makes the second-stage comparison unusually clear.

A read-only menu bar view across Codex, Claude Code and Antigravity makes active sessions visible without changing the runs. Kit tackles the history side of the same problem.

A versioned specs directory makes agent context inspectable: the instruction stays short, and changes to the contract show up in a diff.

Catching mixed timestamp formats in the traces and fixing the contract across tools is good, practical agent work.

Ten agent PRs aren't a tenfold speedup over one human PR. Once agents change the volume, the old PR baseline no longer measures progress. We need to know whether delivery improved. https://hevmind.com/writing/the-software-factory-velocity-trap/?utm_source=bluesky&utm_campaign=velocity-trap

Speed is the wrong target for a software factory. Agent PR volume breaks the velocity baseline, and competitors have agents too. We’d measure quality, reliability and performance instead. https://hevmind.com/writing/the-software-factory-velocity-trap/?utm_source=bluesky&utm_campaign=velocity-trap

The clever bit in this explanation is treating future tokens as the batch a single-user laptop otherwise lacks.

We gave Claude Code and Codex a way to search past work: kit’s hev-query skill tries several phrasings through `hev query`, then reads the session around the best hits. The archive stays in your turbopuffer account. https://hev.dev/kit/?utm_source=bluesky&utm_campaign=kit-tracing

Beau's single-Postgres build is a useful way to teach hybrid search: BM25 and vector retrieval share one corpus, so the moving parts stay visible.

Having dense, sparse, multi-vector and reranking models in one Sentence Transformers suite makes it much easier to compare retrieval stages without changing libraries.

We wouldn't choose a reranker by model size. On our three-corpus test, Clef pair scored 0.494 mean nDCG@10 at $3.69/1,000 queries; Clef-flash pair scored 0.498 at $1.39. https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

We reversed the candidate list. Clef-flash changed its top document on all 20 diagnostic queries per corpus, in batch and choice. Small slices, but enough to warrant an order-sensitivity test. https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

The re:Invent eval track talks about LLM judges and agent outcomes. Only two sessions measure retrieval quality directly. We marked both on Adam's route, along with places to talk through ranking and recall: https://hev.dev/reinvent/?utm_source=bluesky&utm_campaign=reinvent-retrieval

The re:Invent expo floor has turbopuffer and Zilliz booths. Pinecone, Weaviate, Qdrant and LanceDB weren't listed as exhibitors as of Oct 2. We marked the booths on Adam's search route: https://hev.dev/reinvent/?utm_source=bluesky&utm_campaign=reinvent-retrieval

A failed agent run is easier to find by what happened than by its session timestamp. We index prompts, replies and tool calls together in kit, so `hev query` can take you back to the command and its surrounding exchange. https://hev.dev/kit/?utm_source=bluesky&utm_campaign=kit-tracing

Running predicates on compressed columns is a thoughtful design choice: the million-zero example turns a filter into one comparison.

When the next stage needs a relevance cutoff, rank order is not enough. Perplexity choice spreads one unit of probability across the list. Jev batch scores each document. We tested both: https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

The Azure Functions test found where read/write labels stop: the MCP trigger never sends them to the wire. Naming the API role as the enforcing boundary gives builders something concrete to use.

Kit doesn’t wait for a session to end. hevd reads new and growing coding-agent session files, indexing commands and abandoned approaches while the work is underway. The trace stays searchable after the terminal closes. https://hev.dev/kit/?utm_source=bluesky&utm_campaign=kit-tracing

Mining corrections across old agent sessions, then moving the recurring ones into skills, is careful harness work. We built kit to make those traces searchable.

In our reranker test, Perplexity batch matched Jev batch on quality: 0.500 vs 0.501 mean nDCG@10. One HTTP call still billed shared context for each question, making it $11.59 vs $0.54 per 1,000 queries. https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company

Search cost belongs at the task level when an agent makes hundreds of sub-queries. The OpenSearch team describes that workload at re:Invent. We put their session on Adam’s route. https://hev.dev/reinvent/?utm_source=bluesky&utm_campaign=reinvent-retrieval

AWS's re:Invent catalog has S3 Vectors, DynamoDB vector indexes, OpenSearch GPU indexing, Aurora pgvector hybrid search, and ElastiCache semantic caching. The pitch: use the database you already have. We mapped the talks: https://hev.dev/reinvent/?utm_source=bluesky&utm_campaign=reinvent-retrieval

We haven't built project filtering or redaction into kit yet. hevd indexes every coding-agent session on the Mac, including anything the agent saw. Choose a Mac whose sessions you're comfortable storing in your turbopuffer account. https://hev.dev/kit/?utm_source=bluesky&utm_campaign=kit-tracing

We publish kit under Apache-2.0 with the design RFCs and source in one repo. The capture path is worth inspecting before running it: session files can contain secrets, and kit does not redact them yet. https://github.com/hev/kit

Sizing MaxSim blocks against both query and document token counts is a thoughtful fix for memory use on long query batches.

Writing retrieval evals after a prompt tweak broke the pipeline gives the next quiet regression somewhere to show up.

One HTTP call made Clef-flash worse than leaving the BM25 shortlist alone. Scoring documents one at a time worked better. We publish request shape alongside ranking scores for this reason. https://hevmind.com/writing/jev-has-company/?utm_source=bluesky&utm_campaign=jev-has-company