CRTX
DocsSign in →

Business playbooks

CRTX is a horizontal RAG platform, so the fastest path to value is mapping its primitives — collections, permissions, evals — onto a specific business workflow. This page is a set of concrete playbooks. Each names the collection layout, who gets which permission, the retrieval settings to start with, and how to know it's working.

The universal recipe:

One collection per knowledge domain → curate documents → give the right people query or ingest links → watch the eval dashboard → tune retrieval until faithfulness and relevance are where you need them.

#1. Customer support & internal help desk

Problem. Agents (or customers) waste time hunting through product docs, past tickets, and policy PDFs.

  • Collection layout: one collection per product line or per policy domain.
  • Ingest: product manuals (PDF), help-center pages (URLs), policy docs (DOCX), a CSV export of resolved-ticket Q&A.
  • Permissions: query links to all agents; an ingest link to the docs team.
  • Retrieval: similarity to start; enable reranking once volume grows — support questions are precise and benefit from cross-encoder ranking.
  • Win condition: rising faithfulness on the dashboard, falling average handle time. Use the worst-queries view to find the questions your docs don't actually answer, then write those docs.

Deployment note. To put this on a public help site, expose a scoped, read-only entry point rather than user JWTs — see the embeddable widget and API-key design in Programmatic access.

#2. Sales enablement & RFP response

Problem. Reps need instant, accurate answers about the product, pricing tiers, security posture, and competitive positioning — with sources they can trust in front of a customer.

  • Collection layout: "Sales Knowledge" (decks, one-pagers, pricing), "Security & Compliance" (SOC 2, DPAs, questionnaires), "Competitive."
  • Ingest: PPTX decks (slide-per-chunk keeps talk tracks intact), pricing XLSX, compliance PDFs.
  • Permissions: query to the whole revenue org; ingest to sales ops/enablement.
  • Retrieval: threshold (≥0.70) for the compliance collection — you want "I don't have that documented" over a confident wrong answer in an RFP.
  • Win condition: reps self-serve answers with citations; RFP turnaround drops. Every answer's sources are auditable back to the source document.

Problem. Reviewing contracts and answering "what does our standard MSA say about X?" is slow and repetitive.

  • Collection layout: one collection per contract family or matter.
  • Ingest: executed agreements and templates (PDF/DOCX). Multimodal extraction captures tables (fee schedules, SLAs) as structured Markdown.
  • Permissions: tightly held — query to the legal team only, per-matter collections for isolation.
  • Retrieval: mmr — legal docs repeat boilerplate; MMR avoids returning five copies of the same clause. Raise NEIGHBOR_WINDOW so clauses aren't cut mid-sentence.
  • Win condition: faithful, cited answers that link straight to the governing clause. Always keep a human in the loop for legal conclusions — evals are a triage signal, not counsel.

#4. Internal knowledge base / employee onboarding

Problem. Institutional knowledge lives in scattered wikis, Slack threads, and PDFs; new hires can't find anything.

  • Collection layout: "Company Handbook," "Engineering Runbooks," per-team spaces.
  • Ingest: handbook (Markdown/DOCX), runbooks, benefits PDFs, intranet URLs.
  • Permissions: query link to all employees; ingest to each team's owners so their space stays current.
  • Retrieval: similarity with NEIGHBOR_WINDOW=1 for readable, coherent answers.
  • Win condition: "how do I…" questions get answered instantly with a link to the source of truth; fewer interruptions to senior staff.

#5. Research, analysis & due diligence

Problem. Analysts must synthesize large document sets (filings, reports, data-room contents) under time pressure.

  • Collection layout: one collection per deal, company, or research topic.
  • Ingest: 10-Ks and reports (multimodal PDF pulls out the financial tables and chart descriptions), spreadsheets, market-research URLs.
  • Permissions: per-deal collections; analysts share the exact chat transcript (with sources) to partners/decision-makers via chat sharing.
  • Retrieval: higher Top-K for broad synthesis questions; mmr for diversity across a large corpus.
  • Win condition: hours of reading compressed into cited answers; the eval trend confirms the corpus actually supports the conclusions.

#6. Product & operations analytics on documents

Problem. Ops teams need to query structured exports (CSV/XLSX) conversationally without writing SQL.

  • Ingest: CSV/XLSX exports. Each row becomes a chunk carrying its column headers, so "what's the status of order 10432?" retrieves exactly that row.
  • Permissions: query to the ops team; ingest to whoever refreshes the export.
  • Retrieval: threshold for precise row lookups.
  • Win condition: non-technical staff answer their own data questions; re-ingest the refreshed export (deterministic IDs overwrite cleanly).

#Choosing your starting configuration

Your priorityStrategyTop-KRerankingNeighbour window
Balanced defaultsimilarity5off1
Precision / compliancethreshold4on1
Diverse large corporammr6–8(n/a)1–2
Precise fact lookupsimilarity4on0–1
Broad synthesissimilarity/mmr8+on2

Then let the Evaluation dashboard tell you what to change.

#Rolling it out

  1. Pilot with one team and one collection. Prove the quality loop on real questions before expanding.
  2. Assign a curator. RAG quality tracks content quality — give each collection an owner responsible for what's in it.
  3. Standardize permissions. Decide your query vs ingest policy up front.
  4. Automate ingestion for living corpora using the API (see Programmatic access) so the knowledge base never goes stale.
  5. Review the worst-queries list weekly. It's the shortest path to a better knowledge base.

For putting CRTX in front of customers or inside another team's infrastructure, see Self-hosting & deploying in your environment.