Knowledge Sources
SOPs, policies, runbooks, product docs, support playbooks, and uploaded client documents.
Interactive Demo | RAG Systems
RAG Knowledge Base With Evidence Review
Search a sample operations knowledge base, inspect retrieved chunks, and see how a grounded answer is assembled with citations, metadata, and quality checks.
Try The Retrieval Flow
Ask the knowledge base a business question.
Grounded Answer
Vendor invoice exception path
Evidence Review
Retrieved chunks and metadata.
Clients can see which source records shaped the answer. In production, these cards would map back to Postgres metadata and Pinecone vector IDs.
System Architecture
What the demo is proving.
The page is static for portability, but the architecture mirrors the production pattern clients ask for: ingest documents, store metadata, embed chunks, retrieve, rerank, answer, and evaluate.
Chunk + Metadata
Split documents by section, preserve source IDs, owners, update dates, access level, and source type.
Embed + Index
Generate embeddings, write vectors to Pinecone or pgvector, and keep Postgres as the source-of-truth metadata layer.
Retrieve + Rerank
Use semantic search, metadata filters, hybrid lexical ranking, and reranking before sending context to the model.
Generate + Cite
Ask the LLM to answer only from retrieved chunks and attach citations to every material claim.
Evaluate + Review
Check confidence, groundedness, source coverage, and unsupported claims before showing output to the client.
Implementation Pattern
Code-level signals.
The demo keeps the implementation readable so a client can see what the engineering layer would look like underneath the interface.
Chunk payload
{
"chunk_id": "FIN-002.2",
"source_type": "SOP",
"metadata": {
"owner": "Finance Ops",
"updated_at": "2026-06-14",
"access_level": "internal"
},
"text": "Invoices missing a purchase order route..."
}Retrieval pattern
matches = vector_db.query(
embedding=query_embedding,
top_k=6,
filter={"source_type": {"$in": allowed_types}}
)
context = rerank(query, matches)
answer = llm.generate(schema=GroundedAnswer, context=context)