KNOWLEDGE BASE

Documents

Upload, inspect, and monitor content moving through the indexing workflow.

Documents—in this workspace

Ready—searchable now

Chunks—across all profiles

Content library

Newest uploads first

Rows

Document	Status	Source	Profiles	Chunks	Processing time	Uploaded	Updated	Actions
Connect to load documents.

0 documents

Page 1 of 1

RAG WORKBENCH

Search lab

Test grounded answers, strict response formats, hybrid retrieval, and reranking diagnostics.

Retrieved sources

Run a query to inspect the retrieval path.

⌕

No query yet

Your ranked chunks will appear here with retrieval diagnostics.

DOCUMENT DETAIL

Document

Indexed content and vector representations

Profile—

Indexed chunks

Chunk content, generated questions, and stored metadata

# / Page	Section and content	Generated questions	Indexed metadata
Select a document.

Vector indexes

Original, contextual, and synthetic-question vectors stored in pgvector

Vector ID	Chunk	Page	Representation	Source
Select a document.

SERVICE CONFIGURATION

Settings

Tune ingestion, retrieval, and answer generation. Index model and chunk changes require reindexing.

Models

Dense retrieval and second-stage ranking

Embedding model Embedding dimensionsFixed by the current pgvector schema. Reranker model Reranker max length

Reranker max length caps the tokens the cross-encoder reads per (query, chunk) pair — raise it so a long-context model (e.g. bge-reranker-v2-m3) sees a full large chunk; lower it if latency hurts. The header toggle normalizes rerank scores to a model-independent 0–1 scale, keeping the question rerank-score gate portable across reranker models.

Question expansion

Synthetic queries generated during indexing

Generation model

Questions per chunkMinimum rerank score

Retrieval

Candidate breadth and fusion behavior

Candidate KRRF constant

Candidate K controls the pool sent to the reranker. Higher values may improve recall but increase latency. Code grouping collapses results so each class code returns once, with its fragments linked.

Class-code extraction

How doc-pipeline content is ingested

Code section class

Off ingests clean text only — no class-code work. On tags each chunk with its WC class code and description (fed into the contextual embedding). Code section class is the doc-pipeline segment type scoped for extraction (blank = every page).

Upload behavior

Control content-hash deduplication

Allow duplicate uploads creates a new document and indexing workflow for every file upload, regardless of existing document status. Uploads with an External ID continue to update that stable document.

Answer generation

Grounded response model and token budgets

Answer modelThe provider key stays server-side in OPENAI_API_KEY.

Max context charactersMax output tokens

Chunk profiles

Word budgets used by structure-aware partitioning

Service API keys

Keys other services use to call this API. Shown in full only once, at creation.

Name	Prefix	Access	Created	Last used
No API keys yet.