The company
knowledge base
A verified, provenance-tracked record of what your company knows. Your people query it, your agents read from it before they act, and it holds its value when the model or the model vendor changes.
Why it comes first
Almost every company is using AI now1. Far fewer can point to what it changed. The gap is almost always the same thing: the model has never seen the business. It knows the internet. It does not know your pricing, your contracts, your customers, or the decision you made last March and never wrote down.
Grounding AI in what your company actually knows pays off across every use case that comes after it. Skip it, and each new project starts from zero — re-explaining your products, your pricing, your customers, and your rules to a system that has no memory of any of it.
It is also where the evidence points. The single highest reported AI usage of any industry and function pairing was knowledge management in business, legal, and professional services, at 58%2. The firms that sell expertise for a living went here first. That is adoption, and adoption is not payoff; it is also simply where the documents already were. The payoff argument is further down, and it is a cost argument.
Productivity gains from AI are largest in structured, measurable work — 14–15% in customer support, 26% in software development, 50% in marketing output — and shrink in work requiring deeper reasoning3.
A knowledge base is how unstructured company knowledge becomes structured enough to be in the first category.
MIT put the same gap more bluntly: 95% of organizations investing in generative AI are seeing no measurable return4.
How it's built
Five stages, from a raw source to an answer you can check. The middle three are where the work actually is.
Ingest
We connect the systems the business already runs on and pull from them continuously. An export tells you what was true the week it ran. A live connection keeps the knowledge base current with the company.
- Two kinds of source, handled differently. Documents stay connected at their origin.
- Structured records are declared once as a source, a key and a filter, then run on a schedule. Every row carries a _provenance record of source, run_id, fetched_at, transform, and confidence.
- The raw payload is landed before anything is transformed, so a run can be replayed.
- A verify gate that fails closed decides whether the run is allowed to exist.
Structure
Everything lands in three stores at once, because no single one answers every kind of question well. A search box hands back documents. This hands back answers with the relationships intact.
Documents, entities, and their metadata, stored exactly as they are. This is what you audit when someone asks where an answer came from.
Embeddings so a question about "our refund policy" finds the paragraph that describes it without ever using that phrase. Keyword search cannot do this, which is why most internal wikis fail.
Entities and the relationships between them — which client belongs to which contract, which document a fact was drawn from. Every fact card is anchored to a node in the graph, so when a document changes the system can find exactly the cards it touches, and retire any card whose node is gone.
- Records that describe the same thing get resolved by identity, never by string match.
- Bands: EXACT · FUZZY ≥ 0.92 · AMBIGUOUS 0.80–<0.92 · NONE.
- Only exact and fuzzy merge automatically.
- Ambiguous goes to an adjudicator that says no when it is unsure.
Make a fact
Nothing here scrapes facts out of raw text. The system decides which question to ask, then answers that question against retrieved evidence. The discipline holds whether the corpus is a codebase or a filing cabinet.
Can this vendor raise prices without our sign-off?
- Fact card
The claim: Meridian can raise prices unilaterally with notice, no sign-off required. The card holds that sentence and a pointer back to the clause it came from. It never copies the clause itself.
- Passage
The clause retrieval actually returns from that agreement: the pricing section, surfaced because it matches what the question means, whether or not it shares a single word with it.
- Document
One filing inside that corpus: the Meridian Logistics master services agreement, the actual contract governing the relationship, kept connected at its original address so a citation can always lead back to the real page.
- Corpus
Every vendor contract the company has ever signed, connected as one source: gigabytes of PDFs, scans, and amendments no single person has read end to end.
A passage and a fact card are not two different systems. Both live in the same store, and one field is the only thing that tells them apart: whether a row points back to a document or forward from a source passage. That single index is what makes the jump from a claim to its evidence a single lookup instead of a second search.
An agent asking that question reads the fact card first: the claim and a pointer, nothing else. It opens the underlying passage only to confirm the wording, and the source document only if the passage itself looks wrong. Skip that map and the same agent is left facing gigabytes of unindexed filings, with four ways to fail at once: hallucinating an answer, spending enormous token budgets rereading text it has already scanned, answering too slowly to be useful, or giving up before it finds anything at all.
- The question comes first: a structural spine for a code corpus, a curriculum for a document corpus, and the questions your people actually missed.
- Each question is answered against passages pulled for it.
- Four rules hold for every source. Empty retrieval never reaches the model.
- Retrieved passages enter the prompt as untrusted data, and any instruction sitting inside them carries no authority.
- NOT FOUND is a success path the system is allowed to take.
- A card citing a passage that does not exist is dropped before anyone sees it.
Publish gate
A fact does not become an answer because a model produced it. It has to clear a gate first, which is the step most implementations skip.
- Deterministic checks run first, then a metered groundedness judge.
- Every error path fails closed: a checker that errors is a checker that says no.
- Visibility is re-derived on every pass and defaults to closed.
- You do not have to take our word for it. A machine checker walks every citation and confirms it resolves to the passage it claims.
Ask
A plain-language layer over all of it, serving two audiences from one substrate: your people ask it questions, and your agents read from it before they act. A wiki serves humans. A vector index serves machines. This serves both.
…3 key flags for today:
- 1.
Acme Co is in final stages of negotiation, targeting $180k MRR. John on National accounts is on it. Want me to Slack him for an update?
- 2.
Last night's opportunity screener just identified a new 20,000 sq ft building construction permit filed by Secretive LLC in Brooklyn. Our corporate piercer connected the LLC to Innovatech Solutions, Inc and Mike Smith. Should we draft outreach — we're likely 11 months pre-opening, so we can likely be their first outreach.
- 3.
Vertex Global Ltd. hasn't placed a replenishment order in 6 months (vs. average 4-month cycle over the past 5 years). I'd suggest you check in and pitch a discount on our latest SKUs?
Acme Co, Secretive LLC, Innovatech Solutions, Vertex Global and Mike Smith are invented, so the mechanism reads clearly without exposing anyone real. This is a scripted illustration of the interaction, never a captured product answer.
- Each question runs two independent retrievals at once: pgvector finds passages by meaning over an HNSW index, and Postgres full-text search finds them the way a search engine does, by matching the words themselves.
- The two ranked lists are combined by reciprocal rank fusion. A passage ranking respectably on both outscores one that tops a single list and is missing from the other, so agreement between two different retrieval methods beats a strong showing in either one alone.
- On top of that fused score sits a multiplier: recency, source authority, and a handful of other signals scale a passage up or down without a second sorting pass overriding the fusion.
- For the organizations running it today, ours included, a Voyage cross-encoder reranks that fused list once more, reading each passage against the question directly: a slower, pricier pass spent on accuracy. Off by platform default everywhere else, and off for every other organization until it is turned on for them too. If the reranker's own credentials are wrong, the call fails outright instead of silently serving the fused-but-unreranked order anyway; any other fault falls back to that same fused order without stopping the answer.
- A calibrated floor of 0.45 then drops whatever is left too weak to answer with, whether or not that pass reordered it, so a passage never reaches an answer just because nothing better came back.
- The knowledge lane answers with machine-checkable citations, each one resolvable to the exact passage a claim was drawn from.
- The chat lane answers in plain language, grounded on tool results from the same stores. It links identifiers and carries no citation envelope.
- Every agent deployed afterward inherits the same grounding your team has, and answers out of your company.
Nothing here
holds still
Most retrieval systems are a straight line. Documents go in, chunks come out, and an index starts aging the day it is built. Ask one something it cannot answer and nothing happens: the gap sits there until a person notices it. Change the underlying document and nothing happens either, because the answer built from it has no idea the ground moved. We built both of those into the input.
A miss changes nothing. The index only ever ages.
A miss is the input. Use sharpens it. A new source widens it. A changed source refreshes it.
The re-anchor step is what closes the circle, and the sweep is what keeps the circle honest. The knowledge base improves because people used it, improves again because you connected something new, and corrects itself when something you already connected changes underneath it. The questions your team gave up on are the ones it goes after first.
- 01The ask is recorded
Every question put to the knowledge base is logged with whether it found an answer. Question text stays in the staff-gated store, and the analytics event carries counts and buckets only.
- 02A gap files itself
A question that comes back empty, or comes back weak, enters a queue keyed on the question itself. Ask it a second time and the counter moves instead of the row duplicating. Demand sorts itself into priority order without anyone grooming a backlog.
- 03Triage decides who answers it
Each queued question is classified against the sources already connected. Where the material is there, the question is promoted ahead of planned work and answered automatically. Where it is not, the question parks as a source gap, because no generator can invent a document that was never connected.
- 04The answer earns its place
A promoted question runs the same generation path and the same publish gate as everything else. Wanting an answer badly does not buy it a faster route in.
- 05Parked questions wake up
Source gaps are re-checked as new systems connect. Add a source this quarter and questions filed against it months ago promote themselves and get answered, without anyone thinking to ask again.
- 06A changed source restages its own cards
On the corpora running it today, ours included, a daily sweep rebuilds each source and compares it against the version the last sweep recorded. Only the cards whose evidence actually moved are asked again; everything else is left alone, so keeping the base current costs a fraction of building it. A card whose source document has disappeared is retired instead of being left to answer for a document nobody can open.
- 07A moved citation counts as a change
A document can keep every word and still move to a new address, which would leave a published card pointing at a page that no longer opens. The sweep compares each citation target alongside the text itself, so those cards restage on the next pass and the link a reader clicks keeps resolving.
What it returns
The measured return is cost savings, and it can arrive early.
Knowledge management is one of the clearest cost stories in the survey data: less time spent hunting for what someone already wrote down, faster onboarding, and fewer decisions made against stale information.
Then it compounds. Every AI project you run after this one starts against a company its tools can actually read, instead of starting over. That is the case for doing this first.
Retrieval is a technique, and techniques arrive with whatever tool you bought last. What gets built here is an asset: a curated, provenance-tracked record of what is true about your business, built out of sources you already own. It outlasts the model that reads it and the vendor behind that model.
What we need from you
Read access to where the knowledge actually lives. A knowledge base built from a curated folder of tidy documents describes a company that does not exist.
The most valuable knowledge in most companies has never been written down. Interviews are how it gets captured.
One person accountable for decisions and adoption. Engagements without one produce systems that work and nobody uses.
We buy where buying fits, and it usually does. In MIT's interview sample, external partnerships built on customized, learning-capable tools reached deployment about 67% of the time, against about 33% for tools built in-house7. We read that as a caution about our own instinct to build.
Where we do build, we build on Reeve, our own agentic toolkit: the same one running our companies. Price that in as the conflict of interest it is. What makes it safe is where the knowledge lives: your documents stay connected where they already live, and every card points back to a source you control.
Where would we start?
Tell us what you have tried and where it stalled. We will show you what a knowledge base would change for your business, and where we would start.
Sources
- 1Organizational AI adoption reached 88% of surveyed organizations in 2025. Stanford HAI, 2026 AI Index Report, Chapter 4: Economy (Data: McKinsey & Company Survey, 2025). Retrieved 2026-07-27.
- 2The highest reported AI usage of any industry and function pairing was knowledge management in business, legal, and professional services, at 58%. Stanford HAI, 2026 AI Index Report, Chapter 4: Economy (Figure 4.3.3, p.194). Retrieved 2026-07-27.
- 3Productivity gains from AI are largest in structured, measurable work where outputs are easy to monitor: studies report 14% to 15% in customer support, 26% in software development, and 50% in marketing output. Gains are smaller in tasks requiring deeper reasoning. Stanford HAI, 2026 AI Index Report, Chapter 4: Economy (Chapter Highlights, item 9, p. 174). Retrieved 2026-07-27.
- 4MIT found that 95% of organizations investing in generative AI are seeing no measurable return on it. MIT NANDA, The GenAI Divide: State of AI in Business 2025. Linked copy hosted by mlq.ai. Retrieved 2026-08-26.
- 544% of organizations using analytical AI for knowledge management report a decrease in costs. Stanford HAI, 2026 AI Index Report, Chapter 4: Economy (Figure 4.3.4, "Cost decrease and revenue increase from analytical AI use by function, 2025"; data: McKinsey & Company Survey, 2025). Retrieved 2026-07-27.
- 617% report analytical-AI knowledge-management cost reductions of 10% or more. Stanford HAI, 2026 AI Index Report, Chapter 4: Economy (Figure 4.3.4, "Cost decrease and revenue increase from analytical AI use by function, 2025"; data: McKinsey & Company Survey, 2025). Retrieved 2026-07-27.
- 7In MIT NANDA’s interview sample, external partnerships built on learning-capable, customized tools reached deployment about 67% of the time, against about 33% for tools built in-house. MIT NANDA, The GenAI Divide: State of AI in Business 2025 (p. 19; interview sample of 52 organizations). Linked copy hosted by mlq.ai. Retrieved 2026-08-26.