ABP product · BPMDoc
SmartSearcher
Enterprise search and answers over your documents — with sources, without made-up facts.
Ask a question in natural language — the system finds relevant fragments in your corporate corpus, re-ranks them, and builds an answer with citations: not only the document, but the exact page. Clicking a source opens the PDF positioned on that page. Agents handle routing and tool calls where flexibility pays off.
Built by ABP. The SmartSearcher product.
In one environment
- Hybrid search + re-rank
- Answers only from retrieved documents
- Source down to PDF page, open with position
- Out-of-the-box integrations (Confluence and more)
- Local login or Keycloak / SSO
- ACL from data sources
- Pipeline core, agents where they pay off
Product in action
SmartSearcher UI: knowledge sections and search with answers and sources. Browse the screens or click a shot to open it large.
Company knowledge exists. Finding it is hard.
A typical picture that enterprise search and RAG platforms address: knowledge is scattered, public models do not know your environment, and documents must not leave the perimeter.
Documents in many places
Policies, instructions, mail, project materials — employees waste time on manual search or “asking a colleague”.
Public AI is not enough
A cloud model does not know your standards and internal rules and must not receive confidential corporate data.
Answers need evidence
For engineering and regulations, a confident paragraph is not enough — you need an answer grounded in a specific document fragment.
How it works
A full RAG pipeline: from file upload to an answer with links. Each stage is a dedicated step under a single core.
Ingest
Documents land in object storage. Indexing jobs go to a queue.
Index
Parsing, chunking, embeddings, write to vector store and metadata.
Hybrid search
Semantic, lexical, and full-text channels with fusion ranking.
Answer with sources
Re-rank → context assembly → LLM generation. The answer cites document and page; a click opens the PDF on that page. No fragments — an honest refusal.
Research
Pipeline or agent: what your GPU can actually handle
We compared SmartSearcher's fixed retrieval pipeline with agentic RAG on real support tickets over a private knowledge base — with open-weight models on self-hosted hardware.
More concurrent users on one model GPU at equal latency.
Predictable cost per ticket: nearly constant for the pipeline; the agent's costliest ticket was 2.5× the median.
No quality trade-off: neither design wins robustly on quality — checked by three independent LLM judges.
Sheynin, D. (2026). Pipeline or Agentic RAG on a Single Model GPU: Four Times as Many Concurrent Users at Equal Latency and Predictable Per-Ticket Cost. BPMDoc. Zenodo. doi:10.5281/zenodo.23599687 · CC BY 4.0
Cite (BibTeX)
@misc{sheynin2026pipeline,
author = {Sheynin, Dmitry},
title = {Pipeline or Agentic RAG on a Single Model GPU: Four Times as Many Concurrent Users at Equal Latency and Predictable Per-Ticket Cost},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.23599687},
url = {https://doi.org/10.5281/zenodo.23599687}
}Agents — where they pay off
AI agents are a real breakthrough, and SmartSearcher uses them. But even the most capable tool should be applied with thought. The core path — from a question to an answer with sources — runs as a fixed pipeline: it keeps the cost per ticket predictable and lets one model GPU serve more concurrent users at equal latency. Agents take over where flexibility matters more than predictability.
Agents
- Routing requests to topical assistants
- Tool calling during the dialogue
Fixed pipeline
- Retrieval, re-ranking and answer generation for every question
- Nearly constant cost per ticket
- Room for more concurrent users on your own hardware
Not either-or — the right tool for each step. See the research ↑
Capabilities
Grounded answers
Answers are built only from retrieved fragments. No source — no invented fact.
Hybrid search
Dense + sparse + FTS: catches both meaning and exact codes, identifiers, and designations.
Citations to the page
The system points not only to the document, but to the relevant page. Clicking a source opens the PDF positioned on that page — no manual scrolling.
Async processing
Indexing via queue and workers — uploads do not block user dialogue.
Web UI
Search, document management, knowledge sections, indexing statuses.
Customer perimeter
Docker Compose or Kubernetes. Images can be delivered for air-gapped deployment.
Integrations, auth, and access
Enterprise setup: connect corporate systems out of the box, familiar employee login, and permissions aligned with knowledge sources.
Corporate systems out of the box
Integration with corporate information systems — for example, Confluence — without a custom connector build for every pilot. Knowledge sources plug into the product runtime.
User authentication
Sign in locally (SmartSearcher accounts) or via corporate identity — for example, Keycloak / SSO. One access point for customer employees.
Access control and source ACL
Role model and view restrictions on materials. Access control integrates with data sources — for example, Confluence restrictions: users only see in search what they already may see in the source.
Security and deployment
Data stays with you
- Deployed in your internal environment
- Documents in object storage on your infrastructure
- Local auth or Keycloak / corporate IdP
- ACL and inherited restrictions from sources (Confluence, etc.)
- Action logging
Delivery options
- Standalone: Docker Engine + Compose
- Production: Kubernetes, image registry
- LLM in a private scenario
- Scaling of indexing workers
Technology stack
API
Python, Litestar (async)
UI
Next.js, React, TypeScript
Search
Qdrant, PostgreSQL FTS, re-ranker
Queue
RabbitMQ, async workers
Storage
MinIO, PostgreSQL, Redis
Auth & IdP
Local accounts, Keycloak / OIDC
Integrations
Confluence and corporate systems
Platform
Docker, Kubernetes
Product status
SmartSearcher is an ABP product (BPMDoc brand). It is a full-featured enterprise RAG / multi-agent knowledge search solution.
The product is intended for pilots and production in the customer’s environment: corpus ingest, indexing, hybrid search, answers with sources, container deployment.
Discuss a SmartSearcher pilot
Tell us about your knowledge corpus and environment — we will propose a pilot architecture and a rollout plan in your perimeter.
Prefer email? Write to info-abp@bpmdoc.com