03 · embeddings · chunk strategy · rerank · citations
retrieval that answers from your stuff
doc and product corpus retrieval with citations — or it shuts up.
from $16k · 3–5 weeks·one corpus, citations, fail-closed
30 min · no workshop · no deck
// for who
teams whose chatbot invents pricing, invents policy, or pastes half a pdf into a giant prompt and calls it knowledge.
// not for
dump every pdf and hope · pure fine-tune vanity · chat widget with no corpus care
// what you leave with
- one corpus wired with citations and fail-closed modes
- chunk and rank choices you can change without rewriting the app
- silence when confidence is trash — not a confident lie
// what you get
- one corpus designed: sources, chunk strategy, embed + rerank path
- citation UI or silence contract — never invent a page number
- fail-closed modes when retrieval is weak
- basic ACL / role scoping when the brief needs it
- freshness hook so yesterday's policy doesn't answer forever
- eval seed against real questions that hurt
- handoff: indexes, config, run notes
// out
- unlimited connector sprawl under one band
- company-wide knowledge platform as the first engagement
- fine-tune instead of retrieval as the whole job
- marketing chatbot with no product spine
// how it runs
- 01
name the corpus
what truth, who may see it, what wrong looks like.
- 02
slice one answer path
citations or silence on real questions.
- 03
build retrieval + surface
chunk, rank, UI, fail-closed.
- 04
ship + keys
you own indexes and the path. 90 days on breakage.
// the stack
embeddings · chunk strategy · rerank · citations
named tools. no stack theater. if it isn't on the critical path for this line, it isn't listed.
// the band
from $16k · 3–5 weeks
one corpus, citations, fail-closed
bands filter the room. exact number after a call — not a 40-page sow.
// faq
- why not just stuff the docs into the prompt?
- context windows fill. policies rot. you can't cite a blob. retrieval with chunking, rank, and fail-closed modes is how answers stay honest as the corpus grows.
- will it still invent pricing?
- that's the bug we design against. citations required, silence when weak. if your data is garbage, we say so before the invoice — we don't paint over it.
- how is this different from product-ai?
- product-ai is a feature path with model calls. rag is the retrieval layer when the job is answer from your stuff with provenance. many product-ai builds include a thin rag path; full corpus iron is this line.
- how do you price this?
- from $16k · 3–5 weeks for one corpus, citations, fail-closed. exact number after a call.
// adjacent lines
book if you're not sure which. these two sit next to this line.
tell us what you're building.
30 min · your clock · no deck·hello@nightshiftglow.studio · we answer at 3am