Candidate hits are useful, but provisional.
Search results help locate promising pages and chunks. They are leads that need readback before they become answer evidence.
Proof principle
Search is a lead. Fetch is the proof.
The strongest promise is not vector search. It is the movement from a question to ranked evidence, then to fetched source context that can support the answer.
Fetch turns a hit into citation-grade context.
Fetched chunks provide the text, table, figure, or page neighborhood that lets the final answer stay grounded.
The public lane stays document-centered.
Source-code work belongs in repo/GitHub readback. This lane is focused on PDF and document evidence.
Pipeline
A visible path from document to answer.
The chain stays visible from ingestion to citation: material enters the corpus, context is extracted, candidates are ranked, proof is fetched, and the answer can cite what it used.
PDFs and documents enter corpus authority before they are searchable.
Docling and OCR lanes turn raw files into retrievable text, tables, figures, and metadata.
PostgreSQL, pgvector, and metadata support hybrid candidate discovery.
The tool returns ranked candidate evidence with document-aware signals.
The exact proof chunk is retrieved with bounded context.
The answer cites fetched evidence instead of relying on a model guess.
Evidence forms
Documents are not one kind of text.
A serious document evidence system needs to preserve the shape of the evidence: paragraphs, tables, figures, links, and page-level context all have different verification value.
PDF text chunks
Extracted paragraphs with document and page provenance.
Structured evidence
Tabular material remains fetchable when a row or column proves the answer.
Visual context
Figure records can bring the surrounding page context along with the visual evidence.
Document references
Link records preserve navigational evidence and related material.
Neighbor context
Page-level fetches can widen the view when a single chunk is too narrow.
Boundary
Grounded Evidence is deliberately not the memory layer.
Evidence proves claims from documents. Long-term memory preserves orientation and operating history. They cooperate without becoming one vague knowledge box.
Do not store session residue here.
Decisions, runbooks, and reusable operating notes belong in wiki memory, not in the evidence retrieval lane.
Do not cite search snippets as final proof.
A ranked hit helps navigation. A fetched chunk gives the accountable evidence context.
Derived search state does not outrank documents.
Indexes, vectors, and metadata accelerate discovery, but the document evidence remains the claim authority.
The evidence engine sells proof, not memory.
Search narrows the field, fetch opens the source, and the final answer can point to the document context that made it credible.