METHOD NOTE

Useful answers need
bounded evidence.

This assistant answers from a pre-ingested snapshot of curated public DarGlobal and Wasalt material. It does not crawl the web while you wait for an answer.

Current indexed snapshot

Thirteen curated public pages

The corpus includes six DarGlobal overview, portfolio, investor, editorial, and project pages, plus seven Wasalt platform, city, and corporate pages. It is not a claim that every public URL has been indexed. Each crawl stores the source URL, title, timestamp, cleaned Markdown, and content hash before the text is chunked and indexed.

Technical architecture

Ingestion and retrieval flow

LangGraph architecture showing the guardrail, planner, general-conversation, retrieval, evidence-review, query-refinement, answer-author, abstention, and citation-guard nodes.
LangGraph compiles this bounded workflow. The solid lines are the normal path; the dashed gold lines are the blocked or one-pass refinement branches.

Simple single-topic property questions take the fast route. Complex questions can be decomposed into up to five subquestions, reviewed, and rewritten once before an answer is produced. A general greeting takes the isolated conversation route and never searches the knowledge base.

Controlled expansion

What the crawler excludes

Source allowlist

Only the reviewed public DarGlobal and Wasalt hosts above are eligible for this corpus.

Refresh strategy

Re-run the ingestion job on a schedule or after a reviewed seed update; do not crawl during a chat request.

Explicit exclusions

Login and registration flows, forms, privacy pages, duplicate URL variants, personal contact details, and an unbounded crawl of live individual listings.

Dynamic inventory is a timestamped public snapshot, not a complete live catalogue. New seed URLs are reviewed before they are added to the crawler configuration.

Answer lifecycle

  1. Crawl offline. Crawl4AI renders approved public pages and produces cleaned Markdown.
  2. Preserve provenance. Postgres retains the source document and metadata; Qdrant stores retrieval vectors.
  3. Retrieve deliberately. Hybrid BM25 and vector search selects up to eight evidence chunks.
  4. Respond with proof. The answer agent can cite only the retrieved chunk IDs; the citation guard maps them to the source cards shown in chat.

Design decisions

Pre-ingested, not live-crawled

It gives predictable latency, reproducible answers, and an auditable source snapshot.

Grounded by default

Unsupported property facts, pricing, availability, and terms are declined instead of inferred.

Bounded agent workflow

Compound questions may create up to five retrieval subquestions and one focused rewrite; simple requests use the faster single-query path.