Rajesh Pentakota

Building CodePuP AI | ex- Zynga, Flipkart, Walmart

[Rajesh Pentakota @rajpeko1]

Same vector length doesn't mean same map.

Pair index/query models by documentation: Cohere says Embed 5 Pro and Fast share a space, but says nothing about Embed 4.

Before reusing an index, test nearest neighbors on your corpus.
https://t.co/KlAzBOqDsg
https://t.co/8XhzeosdMX

[Rajesh Pentakota @rajpeko1]

Same embedding space makes a model split possible—not better.

Pro can embed documents; Fast can embed queries against those vectors. Same map, different jobs.

That’s compatibility, not a retrieval win. Compare quality, latency, and cost on your corpus.
https://t.co/LlPs5YjcRW

[Rajesh Pentakota @rajpeko1]

Measure quality, latency, and cost for each path. Don’t assume native embeddings win: Cloudflare documents the mechanisms, not a universal quality advantage.

Test your corpus. Same images, same queries; let results choose.
https://t.co/tYYK394G3q
https://t.co/tfiinTa76Y

[Rajesh Pentakota @rajpeko1]

A chart screenshot can contain the answer in its pixels—and lose it in its caption.

To test that, compare two retrieval paths, not two prompts.

[Rajesh Pentakota @rajpeko1]

Make the comparison fair: index the same corpus in separate instances. Cloudflare fixes the embedding model per instance, so this isn’t a per-query toggle.

Keep retrieval settings and OCR matched. Use fixed queries; report screenshots, charts, and product images separately.

[Rajesh Pentakota @rajpeko1]

Retrieval relevance is finding a witness on the topic. Utility asks whether their testimony improves the answer.

Score passages by their contribution to answer quality—not relevance labels alone. A witness can distract; no single metric wins everywhere.
https://t.co/8lnyGWWy5j

[Rajesh Pentakota @rajpeko1]

A citation check against the model-facing summary can pass while the link to the source is weak. Check each answer claim against both the summary and cited original span. Test with a case where compression drops a detail.
https://t.co/xvZJNQscEG

[Rajesh Pentakota @rajpeko1]

AstaBrief isn't faster retrieval: it gives the writer a question and retrieved excerpts, then generates a report without snippet summaries or clustering in Asta's slower pipeline.

Score usefulness and citation support separately.
https://t.co/15ePKNqwbo

[Rajesh Pentakota @rajpeko1]

Build: inspect the checker and evaluator. Reject infeasible candidates; compare objective values only among feasible ones. Log runtime separately. Did it produce a feasible solution, or only improve the number measured?
https://t.co/uVb9eTE4Dx
https://t.co/NpiC0yBbaQ

[Rajesh Pentakota @rajpeko1]

A heuristic can improve its score while failing the task. Toy port schedule: overlap two ships at one berth to cut travel cost. The metric improves; capacity is violated. This is a generic failure mode, not a reported LACE incident.

[Rajesh Pentakota @rajpeko1]

LACE’s paper reports that five LLM baselines produced no feasible algorithm on four structurally new port-logistics problems. That is evidence on those tasks—not a claim of general algorithm discovery.

[Rajesh Pentakota @rajpeko1]

A constraint validator is a bouncer: it checks encoded rules and rejects violations. A separate scorer ranks who gets in.

Passing the door proves legality, not quality. Track feasible-result rate separately from score among feasible results.
https://t.co/zOcoSl8eQP

[Rajesh Pentakota @rajpeko1]

Don’t tune a heuristic portfolio on the instances you use to judge it.

Selection can learn benchmark quirks, not which heuristic generalizes.

Split instances first: tune on development cases, freeze the portfolio, then test on held-out cases.
https://t.co/OKLpwirBT6

[Rajesh Pentakota @rajpeko1]

Best-of-portfolio isn't a solver that picks a heuristic.

It measures coverage: run the portfolio, then count the best result per instance. To compare with one heuristic, match total search and API-call budgets; evaluate runtime selection separately.
https://t.co/LPYzpvrhOO

[Rajesh Pentakota @rajpeko1]

Test the authority handoff, not just whether the model spots a bad idea. Mímir reports retrospective evaluation on historical data, not a field deployment. A bounded command is not a guarantee of crop safety under arbitrary forecast error.
https://t.co/Wh36cokH1U

[Rajesh Pentakota @rajpeko1]

An LLM can propose an irrigation action without owning the actuator.

Mímir’s control path:
LLM proposal → soil-water simulation → bounded candidate selection → runtime assurance → final action gate → actuator

The proposal is input to the controller, not the command.

[Rajesh Pentakota @rajpeko1]

Test the boundary: inject a negative proposal and one above the allowed bound. Log the proposal, simulator feedback, candidates, selected action, assurance/fallback result, and final command. Assert the actuator-facing value stays in [0, Imax].