How it works

From a bucket of PDFs to a cited answer.

KnowledgeV8 is not a pgvector wrapper. Vector search is one narrow step in a wider pipeline — raw storage, curated knowledge, and a hybrid retrieval that reads full context, not fragments.

Common misread: "it's just pgvector." pgvector only ever handles recall — narrowing a large corpus down to a few candidates. It never touches the raw documents, never holds the reviewed knowledge, and never sees what actually goes into an answer.

Ingestion

How your knowledge gets in.

Three tiers, strict rules about what lives where. Binaries never enter git or Postgres — pgvector only ever sees the small, derived index at the end. Hover a stage, or send a document through.

Sources → Raw → Curate → Knowledge → Index (static view — motion reduced)

hover a stage · click send

01 · Sources

S3 / GCS / Git

A read-only object-storage connector (PDF, DOCX, HTML, CSV) or a git repo of already-written markdown.

02 · Raw

Object storage

Original documents stay exactly where they are — your own bucket, or a platform-managed one. This is the system of record.

store: S3 / GCS / MinIO

03 · Curate

Extract + distill

Text is pulled from raw documents and drafted into OKF concepts. Drafts are rank-penalized until a human — or the workspace owner — promotes them.

04 · Knowledge → Index

Git-reviewed, then embedded

Reviewed concepts live in git, small and diffable. Only at the very last step are they chunked and embedded — into a Postgres index that's fully rebuildable.

store: git (OKF) → pgvector

Retrieval

How a question becomes a cited answer.

Two distinct mechanics, not one. pgvector narrows; long-context expansion widens back out before anything reaches the model. Send a question through to watch both happen.

Query → Recall (narrow) → Expand (widen) → Answer (static view — motion reduced)

hover a stage · click send

01 · Query

A question arrives

Over REST (/v1/answer) or as an MCP tool call from an agent.

02 · Recall

pgvector narrows

HNSW cosine ANN over chunk embeddings finds the handful of concepts that might matter. Recall only — it never generates anything.

03 · Expand

Long context widens

Each surviving concept is read in full — body plus every one-hop linked neighbor — instead of the bare chunk that found it.

04 · Answer

Generate + cite

The model answers from whole concepts and cites the concept ids it drew from — traceable back to the exact source.

pgvector decides what matters. Long context is what actually gets read. Confusing the two is why most RAG answers feel like they're guessing from fragments — KnowledgeV8's citations point at whole, reviewed concepts because that's what the model actually saw.