CC-RAGOS

TL;DR

CC-RAGOS (KnowledgeOS internally) is an enterprise-grade, self-hosted, explainable multimodal RAG platform. Where tools like NotebookLM stop at generating answers, CC-RAGOS shows exactly how every answer was produced — the guardrail check, the query expansion, the chunk retrieval, the reranking, the grounding — streamed live over SSE before the response appears. A custom FastAPI orchestration layer (no Dify, no LangFlow) paired with a Next.js visualization layer means the full pipeline is observable, tunable, and self-hostable. It is simultaneously a production knowledge assistant, a support-ops platform, and a teaching and evaluation environment for RAG engineering.

Market Scope

The Opportunity

The enterprise knowledge management and AI support tooling market is at an inflection point in 2025–2026. Enterprises have spent two years deploying RAG chatbots and are now confronting the trust problem: answers are useful, but nobody can verify them. The shift from "does it work?" to "can I trust it?" is a product opportunity that the current SaaS incumbents are structurally unable to serve — because explainability requires owning the pipeline, and SaaS vendors don't let you in.

The Tailwind

• The enterprise trust gap is now the primary blocker. After two years of hallucinations and uncited answers, enterprise buyers are mandating citation-level grounding and pipeline transparency before any RAG deployment goes to production. This is a new procurement requirement that didn't exist 18 months ago.
• Self-hosting became fast and cheap. In 2025–2026, OpenRouter serves LLM, vision, embeddings, and rerank through one gateway. Qdrant Cloud runs on a free tier. Docling handles multimodal ingestion without GPU for most document types. The total cost to self-host a production RAG system dropped by an order of magnitude.
• Per-seat SaaS pricing is breaking enterprise budgets. Intercom AI and Zendesk AI are running $40,000+/year for a support team — with zero pipeline control and no on-prem option. The economic case for self-hosting has never been clearer.

Target Market

The Problem

Enterprise knowledge is everywhere. Trusted answers are nowhere.

Organizations today accumulate knowledge faster than they can make it findable. The result is not an information shortage — it is a retrieval and trust failure on top of an information surplus.

The structural breakdown has two layers:

1. The knowledge is unstructured and siloed. PDFs, slide decks, spreadsheets, diagrams, screenshots, code files, audio recordings — all sitting in separate systems with no unified query layer. When an employee asks "what is our refund policy for enterprise contracts?", the answer might be buried in a PDF appendix, referenced in a slide, and contradicted by a later spreadsheet update. No search tool surfaces all three. No tool reconciles them.

2. When AI answers do exist, they can't be trusted. Existing RAG tools produce answers but hide everything about how they got there. Users can't see which chunks were retrieved, can't verify citations against source images or diagrams, can't measure whether the answer is grounded in the document or hallucinated around it. For regulated industries, that opacity alone is a procurement blocker. For everyone else, it means the answer gets checked manually anyway — eliminating the productivity gain.

The symptoms teams feel:

• Support agents spending hours on tickets that existing docs already answer — because the docs aren't findable
• "Knowledge gaps" never surfaced — the docs team doesn't know what's missing until a ticket pattern repeats fifty times
• Answers citing images or diagrams with no visual grounding — users can't verify what the AI is pointing at
• Data leaving premises for SaaS processing — a hard no for legal, finance, and healthcare clients
• No way to measure answer quality or compare retrieval strategies over real production conversations
• RAG engineers unable to diagnose retrieval failures because every vendor tool hides the internals

Personas

Primary — The Enterprise Knowledge Worker

"I need a single place to ask questions across all our docs and actually trust the answer — not just hope it's right."

• Who: PM, Sales lead, HR manager, or engineer at a company with a large and fragmented document library
• Context: Constantly switching between PDFs, Notion, Confluence, slide decks, and internal wikis to answer questions that should take 30 seconds
• Pain: When they finally find an AI tool that surfaces answers, there are no clickable citations. They can't verify whether the answer came from the contract appendix or the chatbot's imagination. They end up cross-checking manually — back to square one.
• Goal: Ask once, get a cited answer, click through to the exact paragraph or diagram region that justified it. Move on.
• Frustration quote: "It gave me a confident answer with no source. I spent 20 minutes verifying it manually. At that point I may as well not have asked."

Secondary — The RAG Engineer / Trainee

"I'm trying to learn how RAG actually works — or tune a pipeline that's already in production. Every tool I find hides exactly the part I need to see."

• Who: ML engineer, backend developer, or bootcamp graduate learning RAG systems; or a senior engineer responsible for a deployed RAG product who needs to diagnose retrieval failures
• Context: Reading papers about chunking strategies, HyDE, reranking, and hybrid retrieval — but no tool lets them compare strategies side-by-side over real documents
• Pain: Existing RAG platforms (LangFlow, Dify, hosted APIs) abstract away the internals. You can change a config value but you can't watch what actually happens to a query — which chunks were recalled, what the embedding distance was, how the reranker reordered them.
• Goal: A system where every retrieval decision is observable, comparable, and explainable. A real eval harness that runs against production conversations, not synthetic golden sets.
• Frustration quote: "I changed the chunking strategy and the answers got worse. I have no idea why. The tool doesn't show me what changed."

Tertiary — The Support Ops Manager

"My team is handling the same questions on repeat. I know the answers are in the docs somewhere. I'm paying $40K a year for a tool that can't tell me which questions it deflected or what docs it's missing."

• Who: Head of Customer Support or VP of Customer Success at a B2B SaaS company, managing a team of 5–50 agents
• Context: Running Intercom or Zendesk with an AI add-on bolted on top. High ticket volume, high handle time, no visibility into whether the AI is actually helping.
• Pain: No deflection metrics. No AI draft replies. No knowledge gap detection. Per-seat pricing that scales painfully with headcount. No way to know which docs to write next to reduce repeat tickets.
• Goal: A measurable reduction in ticket volume, a drop in average handle time, and a clear backlog of what docs to write next — all from a single dashboard.
• Frustration quote: "We're paying per seat for a black box. I can't prove to my CFO it's working, and I can't tell my docs team what to write."

The Solution

CC-RAGOS is the explainable knowledge layer that sits between enterprise documents and the teams that need to act on them. It does three things no other platform does together:

1. Ingest Everything

Docling (with PyMuPDF and vision fallbacks) parses PDFs, images, tables, diagrams, audio recordings, and web documentation into a unified knowledge store. Six chunking strategies — including an agentic LLM-detected boundary strategy that identifies natural section breaks and generates chunk titles — let teams tune ingestion to document type. Every chunk lands in Qdrant with both dense embeddings (text-embedding-3-large, 3072-d) and BM25 sparse vectors, fused via Reciprocal Rank Fusion for hybrid retrieval. Cohere Rerank sits on top of the recalled set. The result: any document type, any structure, becomes a searchable, queryable AI knowledge source without preprocessing work from the team.

2. Explain Every Answer

NotebookLM-style grounded answers with clickable source citations — but taken further than any existing tool. For images and diagrams, CC-RAGOS returns the visual bounding box of the cited region and highlights it in the UI: you don't just see "source: diagram on page 4," you see the exact part of the diagram the answer drew from. The Chunk Explorer shows every retrieved chunk and its score. The Embedding UMAP viewer shows the semantic space of the workspace. The Knowledge Graph explorer shows entity and concept relationships across documents. And the flagship USP — "Explain the Pipeline" — streams every retrieval step live over SSE before the answer appears: the guardrail check, the HyDE query expansion, the chunk recall, the Cohere reranking, the grounding step. Users see how the answer was produced, not just what it says. This is the capability that converts skeptical enterprise buyers.

3. Close the Support Loop

An embeddable chat widget (one script tag, zero-ops deployment) brings CC-RAGOS to any customer-facing portal. The multi-tenant Support Desk consolidates queries across portals with a unified agent view. Ticket deflection analytics show exactly how many incoming questions were resolved without agent intervention. AI draft replies with tone refinement let agents respond faster to the tickets that do land. The Knowledge Gap panel surfaces questions that the system couldn't answer confidently — ranked by frequency and routed directly to the docs team as a prioritized writing backlog. A weekly digest report and an MCP server at /api/mcp let support ops managers query deflection metrics and gap trends from Claude Desktop or any MCP-compatible client without opening a dashboard.

Why Now

Three forces converged in 2025–2026 that make this the right moment — and make waiting costly:

RAG tooling matured enough to self-host in a weekend

In 2024, building a production multimodal RAG system required stitching together five separate vendor contracts, a GPU for ingestion, and weeks of integration work. By 2026, OpenRouter serves LLM, vision, embeddings, and rerank through one gateway with model-swappable routing. Qdrant Cloud runs on a free tier with native binary quantization and HNSW indexing. Docling handles PDFs, images, and tables without GPU for the vast majority of enterprise documents. Deepgram Aura-2 adds audio transcription via a single API call. The infrastructure cost to run CC-RAGOS today is negligible. The same system 18 months ago cost 10x more to operate and 3x longer to build.

Enterprise trust is now the procurement gate

After two years of RAG deployments hallucinating confidently, enterprise buyers have changed their requirements. Explainability and citation-level grounding are now evaluation criteria in RFPs, not nice-to-haves. The question is no longer "does it answer questions?" but "can I audit every answer?" CC-RAGOS was designed from day one on the premise that the pipeline must be fully observable — not because it's a clever differentiator, but because it's the only way to build trust in an AI system that touches regulated knowledge. The market arrived at the same conclusion in 2025.

Per-seat SaaS pricing is no longer defensible

Intercom AI and Zendesk AI are charging $40,000+ per year for a support team with no on-prem option, no pipeline visibility, and no knowledge gap tooling. Enterprise buyers are running the math: self-hosted CC-RAGOS costs ~$0.001 per answer in serving compute, with ingestion staying local. The ROI conversation has shifted from "is AI worth it?" to "why are we paying SaaS margins for something we can run ourselves?" The procurement window for self-hosted alternatives is open now — before SaaS incumbents ship on-prem editions and close it.

Differentiation

CC-RAGOS doesn't compete on the ability to answer questions. Every tool does that. It competes on the ability to prove the answer is right.

The moat: three things no competitor has combined

Explainable Pipeline. The "Explain the Pipeline" USP streams every retrieval step over SSE before the answer appears. The user watches the guardrail check fire, the HyDE query expand, the chunks recall and score, the Cohere reranker reorder, the grounding step confirm. No other production RAG tool exposes this. It is the feature that converts regulated-industry buyers who can't take "trust us" for an answer — and it is what makes CC-RAGOS a training platform for RAG engineers, not just a knowledge assistant.

Visual Citations. When an answer draws from an image or diagram region, CC-RAGOS returns the bounding box of the cited area and renders it highlighted in the UI. The user sees exactly which part of the diagram the answer is referencing. NotebookLM, the closest reference product, cannot produce visual citations at all. For any enterprise working with technical documentation, architecture diagrams, or compliance charts, this is a hard requirement that no other self-hostable tool meets.

Own Eval Harness. CC-RAGOS does not use RAGAS or DeepEval off the shelf. The eval harness is a custom LLM-judge plus deterministic retrieval metrics that run against real production conversations — not synthetic golden sets. Production results: faithfulness 0.96–1.00, answer_relevancy 1.000, context_relevance 1.000, citation_recall 1.000. Cost per eval run: ~$0.001 per answer. This means every pipeline change is validated against actual user queries before it ships, and the quality bar is measured continuously in production — not once in a benchmark lab.

Product Architecture

CC-RAGOS runs on a custom FastAPI orchestration layer with SSE streaming, a Next.js 15 visualization and workspace frontend, and Qdrant as the vector store. Fully self-hostable via Docker; optionally deployable with Vercel + Qdrant Cloud free tier for zero-ops teams.

Core Modules

Tech Stack

• Orchestrator: Custom FastAPI (SSE streaming, async) + LangGraph (agentic multi-hop DeepAgent)
• Models: OpenRouter gateway — LLM, vision, embeddings, and rerank swappable without code changes
• Embeddings: text-embedding-3-large (3072-d dense) + BM25 sparse, fused via Reciprocal Rank Fusion (hybrid retrieval)
• Rerank: Cohere Rerank v3 (command-r-plus family)
• Vector DB: Qdrant (native binary quantization, HNSW indexing)
• Graph: NetworkX in-process per workspace (Knowledge Graph, GraphRAG strategy)
• Ingestion: Docling + PyMuPDF + vision model fallback; Deepgram Aura-2 for audio transcription
• Frontend: Next.js 15 (App Router), Tailwind CSS, ApexCharts (dark-themed)
• Auth: Self-hosted JWT — bcrypt password hashing, HS256 tokens, role-based middleware
• Observability: Langfuse (query traces, token cost per conversation)
• Deploy: Docker Compose (full self-hosted) or Vercel + Qdrant Cloud (free-tier cloud path)

Traction & Roadmap

What's shipped (as of September 2026)

CC-RAGOS is live in production serving the Unify and Simplified customer portals at Codeclouds IT Solutions.

Shipped in 2026:

• Agentic RAG via DeepAgent (LangGraph, multi-hop reasoning, sub-agents + scratchpad for complex queries)
• Embeddable chat widget (widget.js) + agent-assist portal overlay (agent-assist.js)
• Multi-tenant Support Desk (Unify portal + Simplified portal, unified agent view)
• Knowledge Graph explorer (per-workspace entity and concept graph)
• Embedding UMAP viewer (semantic space visualization)
• Chunk Explorer (per-query retrieved chunk inspection with scores)
• Visual citations (image/diagram bounding box highlighting in PDF viewer)
• Eval harness: faithfulness 0.96–1.00, answer_relevancy 1.000, context_relevance 1.000, citation_recall 1.000 — over real production conversations
• Audio Overview feature (Deepgram Aura-2 TTS for document summaries)
• MCP server at /api/mcp (read-only reports and deflection analytics tools)
• Auth + RBAC (JWT, three roles — viewer/editor/admin, per-workspace enforcement)
• Agentic chunking (LLM-detected section boundaries + auto-generated chunk titles)
• "Explain the Pipeline" SSE stream (live retrieval step narration before answer delivery)
• Knowledge gap panel (failed-query surfacing + prioritized docs backlog)
• Weekly digest report for support ops managers

Roadmap (Now → Next → Later)

Now

• Provider pinning for main completion model — reduce first-token latency spread from 3–12 s to a consistent 3–5 s band across OpenRouter-routed providers

Next

• ColPali visual embeddings — native vision-language embedding for image-heavy documents (GPU-dependent; queued behind hardware availability)
• Postgres migration — move from in-process storage to Postgres for multi-user production deployments with proper transactional guarantees

Later

• Multi-tenant public SaaS offering — managed cloud path for support teams wanting zero-ops deployment
• OIDC / Keycloak SSO — enterprise single sign-on for organizations requiring federated identity management

Build philosophy

Every eval run in the Part C eval harness uses real production conversations from Unify and Simplified portals — not synthetic golden sets. The rule is simple: fix the measurement before fixing the model. If the eval harness can't detect a retrieval failure, the failure will recur silently. Quality is a measurement problem before it is an engineering problem.

Business Model

Current

Internal platform at Codeclouds IT Solutions. CC-RAGOS is live in production serving the Unify and Simplified customer portals. The serving cost today is approximately $0.001 per answer — ingestion compute stays local; query-side runs on Qdrant Cloud free tier and OpenRouter pay-per-token. Infrastructure cost at current volume is negligible.

Planned monetisation

Why this model works

• Self-hosting eliminates the per-seat SaaS fee immediately. The ROI story is concrete and measurable from week 1: tickets deflected × average handle time saved + AI draft replies reducing per-ticket handle time + repeat tickets eliminated by the gap panel. These numbers come out of the Support Desk analytics dashboard — no custom instrumentation required.
• The license model fits regulated buyers exactly. Legal, finance, and healthcare clients have a hard requirement for data residency. A self-hosted license with setup support is the only model that clears procurement. SaaS competitors are structurally excluded from this segment.
• Per-workspace pricing aligns cost with value. Teams pay for the knowledge bases they maintain, not the number of people who query them. This is a fundamentally different unit economics story from per-seat tools — and a much easier conversation with a CFO comparing it to $40K/year Intercom AI.
• The data asset compounds. The longer CC-RAGOS runs in a workspace, the richer the knowledge gap backlog, the more refined the eval baseline, and the higher the deflection rate. Churn is expensive for the customer — structural stickiness is built in.

Revenue adjacencies

• Eval-as-a-Service: The Part C eval harness — LLM-judge + retrieval metrics over production conversations — is a standalone capability that enterprise RAG teams need independently of the rest of the platform. Exportable eval reports as a premium add-on.
• MCP marketplace distribution: As Claude Desktop, Windsurf, and Cursor mature their MCP ecosystems, the CC-RAGOS MCP server becomes a distribution channel into enterprise AI tooling stacks without a direct sales motion.
• RAG training and certification: The "Explain the Pipeline" feature and Chunk/UMAP explorers make CC-RAGOS a natural teaching environment for RAG engineering. Structured training programs for engineering teams are a natural adjacency once the platform is in wider use.