How it works under the hood
The full TENGWAR architecture: how your documents are processed, how every question is answered, what the operator controls, and the technology it is built on.
From Document to Knowledge
Every step traceable. Every chunk tied back to its source. Section-aware. Multilingual. Audit-ready.
Parsing
Layout, text, assets and fusion, the document is read and structured.
- 1Structure detectionDocling
- 2Text extractionDocling + OCR model
- 3Asset extractionOCR model + MinIO
- 4Content fusion-
Structure & Translation
Section hierarchy, cross-references and language alignment.
- 5Context enrichmentLLM
- 6Structure extractionDocling
- 7TranslationTranslation model
- 8Asset description translationTranslation model
Chunking & Vectorization
Globally optimal boundaries, hybrid index, security-stamped storage.
- 9ChunkingSemantic chunking
- 10Contextual retrievalLLM
- 11Asset linking-
- 12Structure linking-
- 13TokenizationTokenizer model
- 14EmbeddingEmbedding model
- 15StoringQdrant
- 16Summary generationLLM + Qdrant
- 17Acceptance-
Restart-safe
Heartbeat from the ingester to the backend lets the document workflow state machine detect stuck documents and recover them.
Security baked into the index
Clearance level and department IDs stamped onto every Qdrant point at storage time. The query layer participates in the search, not on top of it.
Structure detection
Structural analysis with Docling: section hierarchy, tables, and cross-references are preserved across all 18 supported formats, native or scanned.
Semantic OCR
For scanned documents we use a state-of-the-art vision-language model instead of classic OCR, capturing meaning and context, not just characters.
Figures and tables
Every figure and table is extracted as an image and described with type-specific prompts, so visual content becomes retrievable through its description.
Structural translation
Per-element translation preserves the original document's structure; content stays aligned across languages without collapsing the hierarchy.
Semantic chunking
Instead of splitting by paragraphs, fixed length, or sliding windows, we cluster content by semantic similarity so each fragment represents a coherent idea; section-aware, so fragments never cross headings.
Contextual retrieval
Each fragment can be enriched with LLM-generated context before indexing, improving relevance when retrieved standalone.
Hybrid index
Each fragment is indexed in parallel with a state-of-the-art embedding model for meaning and with BM25 for exact lexical matching. Searches combine both signals.
Hierarchical retrieval
Searches can filter at the document level first, then section, then fragment, using optional summaries at each level to speed up and focus queries.
Inside the Engine
How a question becomes an answer, every step traceable, every claim cited.
Retrieval and answering flow
- User queryWhat the user types in the chat.
- Smart routingThe system reads the question and picks the right path, most questions never need an LLM.
- Question preparationThe system clarifies pronouns, translates if needed, identifies the intent and rewrites the query into useful variants.
- Multi-channel searchFour searches run in parallel against your documents, every angle covered at the same time.
- Smart re-rankingA re-ranker model re-orders the candidates by how well each one really answers the original question.
- Self-correctionThree independent checks catch low-confidence or empty answers and re-run the search with a corrected query before responding.
- Context preparationPicks the highest-value fragments, places them where the LLM reads them best, and attaches the citations.
- One agent, many toolsA single agent grounds the answer in the retrieved context and calls tools, web search, document generation, when the question needs them.
- Agent responseStreamed back token-by-token, with every citation linked to the exact source document and page.
Two-path routing
Simple, unambiguous questions take the fast deterministic path with zero LLM cost. Only the genuinely complex ones spin up the full reasoning pipeline.
Hybrid search
Combines semantic search ("what does it mean?"), classic keyword search ("these exact words") and a knowledge-graph walk ("how does this connect?"), and weights them based on what kind of question you asked.
Three-layer self-correction
If the system can't find a confident answer, it re-runs retrieval with a corrected query before responding. It catches its own bad answers before you do.
Graph-powered search
The fourth channel walks the knowledge graph: it connects entities across documents and brings back facts that neither semantic nor keyword search can reach on their own.
Multi-part questions
Ask several things at once and the system splits them, answers each independently, and tells the LLM how confident it is in each piece.
Live activity & citations
You see what the agent is doing, which tool it is calling and which documents it is citing, in real time, while the answer is still forming.
Security inside the search
Permission checks happen inside the search query itself, not on top of it. Users only see fragments their clearance and department allow.
Full audit trail
Every tool call, every search, every answer, logged with user, department, status and timing. Filterable for compliance reviews and incident response.
Jobs that outlive the chat
Deep investigations, bulk analysis, scheduled digests and document generation run as durable jobs with their own inbox: enqueue, disconnect, and collect the results when they're done.
Your Documents Become a Knowledge Graph
TENGWAR distills your corpus into verified facts, who, what, and how they relate, and turns them into a graph that answers questions plain search can't.
- 1
Fact extraction
During the processing window, the chat model reads every chunk of the corpus and extracts subject–relation–object facts with a confidence level. A support check discards facts the text doesn't actually back.
- 2
Ontology promotion
Entity and relation types aren't hand-defined: they are promoted from the corpus itself into a stable ontology that every new fact is normalized against.
- 3
Conflicts & bitemporality
Contradicting facts are detected and resolved with their evidence. Every fact records when it became true and when it stopped being true, bitemporal validity, and nothing is silently overwritten.
- 4
Graph projection
Entities and facts live in relational tables, no graph database, and are projected into an in-memory, permission-maskable snapshot at query time. Each fact's vectors are indexed in Qdrant.
- 5
PPR retrieval channel
On every question, a Personalized PageRank walk runs in parallel with dense and lexical search; fusion weights each channel by the question's intent, with the same ACL applied to the facts.
- 6
Cited answers
Every fact keeps the chunk and document it came from, so graph-backed answers cite the exact source, just like every other channel.
One store, two spaces: corpus facts and per-user memory share the same engine, separated by design.
Facts, not fragments
Every entry in the graph is a subject–relation–object fact extracted by the AI and tied back to the exact passage it came from.
A fourth search channel
Alongside hierarchy, meaning and keywords, a graph walk connects entities across documents, answering questions no single fragment contains.
Knows when facts change
Each fact records when it became true and when it stopped being true, so answers never mix yesterday's reality with today's.
Contradictions get resolved
When two documents disagree, the conflict engine surfaces both facts with their evidence and records the resolution, nothing is silently overwritten.
Explore it visually
An interactive graph of the whole corpus: filter by entity type, zoom from the big picture to a single fact, and jump from any claim to its source document.
Permissions carry over
The graph enforces the same department-level access control as every other channel, nobody reaches facts from documents they can't open.
Adapts to You, Privately
TENGWAR adapts to each user without sending anything to a third party. The implicit (what it learns) and the explicit (your personal prompt), both feeding the same private workspace.
Two ways to adapt
TENGWAR learns from how you work, and you can also tell it directly with a personal prompt. Both stay inside your private workspace.
A knowledge graph of your own
What TENGWAR learns about you is stored as structured facts, each weighted by confidence, forming a personal graph you can explore visually at any time.
You confirm what it learns
New facts arrive as candidates you confirm or dismiss. When two facts contradict each other, TENGWAR shows both with their evidence and you decide which one wins.
You stay in control
Edit or delete any fact, backed by a full decision history. Erase everything in one action, with a 24-hour window to undo it.
Your prompt is yours
The instructions you give TENGWAR about yourself are never logged in telemetry, redacted at the source.
Strict isolation
Your personal knowledge is yours alone. By design, no other user's session can see it, search it, or leak into your context.
Admin & Operations
- Turnkey doesn't mean opaque. Here's what the operator runs.
- Every agent action, every tool call, every login, auditable.
- Tune the agents. Schedule the ingestion. Watch the GPUs.
Technology Stack
Built with enterprise-grade technologies for reliability and performance.
Agent Framework
Qwen3.6-35B-A3B
GLM-OCR 0.9B
Docling
Enterprise AI Without Compromise
See how TENGWAR works with your documents. We'll set up a private demo with your real use case.