Project: Building a Company Document Bot: Paperless-ngx + Ollama + RAG in Your Own Chat

IRC-Mania Die Seite rund um Sicherheit und Programmierung

Internet, Sicherheit , Internet Relay Chat

Gesponsert

What This Project Does

Employees ask in chat: “What’s the notice period in the contract with Customer X?” or “Where’s the manual for the machine?” The bot searches the company archive and responds with sources. Completely local: no documents leave the office.

The Stack:

Paper/Email/PDF
    └── Paperless-ngx (Scan + OCR + Archive)
            └── RAG Pipeline (Embeddings + Qdrant)
                    └── Ollama (Answer Model)
                            └── Chat Frontend (Mattermost / Nextcloud Talk / Web)

Prerequisites: A mini PC or server (32 GB RAM is enough), the tools from the Software Stack article, about 2-3 hours for setup.

Why This Architecture

• Paperless-ngx is the archive backbone: OCR makes scans searchable, tags and correspondents organize content. See Paperless-ngx.
• RAG instead of training: Documents aren’t baked into the model; they’re searched via embeddings. New documents appear instantly, no retraining needed.
• Chat as the frontend: Employees ask where they already are: Mattermost or Nextcloud Talk. No new tool to learn.
• GDPR-compliant: Everything on-premise. Client and contract data stays internal.

Step 1: Paperless-ngx as Archive

# docker-compose.yml (core)
services:
  paperless:
    image: ghcr.io/paperless-ngx/paperless-ngx:latest
    ports: ["127.0.0.1:8000:8000"]
    environment:
      - PAPERLESS_OCR_LANGUAGE=deu
      - PAPERLESS_TIME_ZONE=Europe/Berlin
    volumes:
      - paperless_data:/usr/src/paperless/data
      - ./consume:/usr/src/paperless/consume   # Drop files here → OCR + Archive

volumes:
  paperless_data:

Drop documents into ./consume (or point your scanner directly there): Paperless handles OCR and archival. See Paperless-ngx Integration.

Step 2: RAG Pipeline: Indexing Documents for Search

# ingest.py, Index Paperless documents into Qdrant
import requests, uuid
from qdrant_client import QdrantClient
from qdrant_client.models import PointStruct

qdrant = QdrantClient("localhost", port=6333)

def embed(text):
    r = requests.post("http://localhost:11434/api/embeddings",
                      json={"model": "nomic-embed-text", "prompt": text})
    return r.json()["embedding"]

# Paperless API: fetch documents
docs = requests.get("http://localhost:8000/api/documents/",
                    headers={"Authorization": "Token YOUR_TOKEN"}).json()["results"]

for doc in docs:
    text = doc.get("content", "")[:8000]          # OCR text
    vec = embed(text)
    qdrant.upsert("firma", [PointStruct(
        id=str(uuid.uuid4()), vector=vec,
        payload={"title": doc["title"], "id": doc["id"], "text": text[:2000]}
    )])

Use nomic-embed-text as the embedding model (ollama pull nomic-embed-text), Qdrant as the vector database. See Local RAG.

Step 3: The Bot: Question → Retrieval → Answer

# bot.py, Core loop
import requests

def answer(question):
    # 1. Find similar documents
    qvec = requests.post("http://localhost:11434/api/embeddings",
                         json={"model": "nomic-embed-text", "prompt": question}
                         ).json()["embedding"]
    hits = qdrant.search("firma", query_vector=qvec, limit=5)

    # 2. Build context
    context = "\n\n".join(h.payload["text"] for h in hits)
    sources = [h.payload["title"] for h in hits]

    # 3. Generate answer, ONLY from context
    prompt = f"""Answer the question ONLY from these documents.
Documents:
{context}

Question: {question}
If the answer is not in the documents: say so honestly."""

    r = requests.post("http://localhost:11434/api/generate",
                      json={"model": "qwen2.5:14b", "prompt": prompt, "stream": False})
    return r.json()["response"], sources

The key: “ONLY from these documents” in the prompt prevents hallucinations. The bot cites the archive instead of guessing.

Step 4: Connect Your Chat Frontend

The bot answers where your company already chats. Examples in the platform articles:

• Mattermost (Article): WebSocket bot, /knowledge question..., self-hosted, GDPR-clean.
• Nextcloud Talk (Article): Webhook bot, built right into your existing Nextcloud.
• Simple web frontend: Open WebUI with RAG support, fastest path, no custom bot code needed.
• Telegram (Article): For mobile too.

Step 5: Operations and Automatic Indexing

A cron job or Paperless webhook triggers ingest.py each time a document arrives. The archive grows, the bot learns automatically:

# Paperless: post-consume-script or simple cron
*/10 * * * * docker exec paperless python3 /scripts/ingest.py --new-only

Extensions

• Per-department permissions: Separate Qdrant collections by department. Sales doesn’t see HR documents.
• n8n variant: Same flow as an n8n workflow, visual, no Python.
• Answer quality: Add a reranking model after embedding retrieval. See Reranking Models.
• Source links: Bot posts direct Paperless links to documents, not just titles.
• Multilingual: Foreign contracts. Choose a multilingual embedding model.

What You’ll Learn

• RAG architecture in real production (not just theory)
• OCR + archive + search as a pipeline
• Hallucination prevention through “context only” prompts
• Chat bots in business context (permissions, GDPR)

Further Reading

• IRC-Coding.de: Programming tutorials on RAG, bot code.
• Local RAG: The technology in depth.
• Paperless-ngx: Archive setup.
• Mattermost Bots: Chat frontend.
• Internal Knowledge Bot: Related scenario.
• Homelab: The hardware.

Key Takeaways:

• Company document bot = Paperless (archive/OCR) + Qdrant (vector search) + Ollama (answers) + chat frontend.
• “Answer only from documents” prompt prevents hallucinations. Core of the design.
• Fully local means GDPR-clean for contracts and client data.
• Indexing via cron or webhook: archive grows, bot learns with it.
• About 2-3 hours to set up on existing hardware.

FAQ

Sources and Further Reading

• Paperless-ngx: GitHub.
• Qdrant: Vector database.
• IRC-Coding.de: Programming tutorials.

Premium VPS Hosting

Schnelle und sichere VPS Server

Gesponsert