Content scattered across formats
PDFs, Word files, scanned pages, images, JSON, Markdown and spreadsheets each need different handling before they can be searched together.
KNOWLEDGE GRAPH RAG PLATFORM
Answerra
From your files to cited, grounded answers
Ingest documents and spreadsheets into a Neo4j knowledge graph and a Qdrant vector index. Ask questions in plain language and get answers from Claude or Ollama, backed by dense, keyword and graph retrieval with checked citations.
The problem
Your knowledge lives in files. Answers need evidence.
Teams hold answers across documents, scans and spreadsheets. The hard part is finding the right passage, connecting it to related facts, and trusting what comes back.
One search method misses things
Semantic search, exact keyword matches and relationships between entities each find different evidence. Relying on only one leaves gaps.
Tables are not paragraphs
Counts, filters, trends and comparisons over CSV or Excel data call for structured queries, not text snippets.
Answers without sources
A generated answer is hard to trust when it cannot point back to the exact passages it relied on.
Conversations lose context
Follow-up questions, corrections and reformat requests break down when the assistant forgets the previous turn.
Unclear cost and graph quality
Without visibility into token usage, orphan or duplicate nodes, and provenance gaps, it is hard to know what the system is doing.
The Answerra solution
Ingest. Connect. Answer.
Answerra builds a document/entity graph, a vector index and optional table store from your files, then combines graph, vector and keyword retrieval to answer questions with citations.
1
Upload
Add files in the browser, CLI or through connectors
2
Parse
Extract text, tables and OCR from scans and images
3
Extract
Sentence-aware chunks, entities, concepts and typed relations
4
Index
Write to Neo4j graph, Qdrant vectors and PostgreSQL tables
5
Retrieve
Dense + BM25 + graph search, or SQL for table questions
6
Rank
RRF fusion, MaxP reranking and evidence selection
7
Answer
Claude or Ollama answer with validated [S#] citations
Ingestion makes no generative LLM calls: parsing, spaCy extraction and local embeddings build the index. A document answer normally uses one generative call.
Features & capabilities
Built for answers you can trace
Capabilities below reflect implemented Answerra functionality — not a wishlist. Optional features are switched on by configuration.
Ingestion & knowledge construction
Broad file support
PDF, DOCX, TXT, Markdown, JSON, PNG, JPG/JPEG, CSV, TSV and XLSX. DOCX keeps paragraph and table order; spreadsheets keep sheet and row-range context.
Benefit: bring the files you already have.
OCR fallback
Scanned or low-text PDF pages, image-only DOCX content and supported image files are read with OCR.
Benefit: scanned documents become searchable too.
Zero generative LLM ingest
An LLM guard ensures parsing, spaCy extraction and local embeddings build the index without calling an answer model.
Benefit: indexing runs locally without model spend.
Entity & fact graph
Documents, chunks, mentions, typed entities and fact edges are written with provenance, plus typed record nodes and reference edges.
Benefit: every fact links back to where it came from.
Resumable, content-aware ingest
Checkpoints, failed-document logs, batch controls and force reprocessing. Completed documents can be skipped; identity is content aware.
Benefit: large libraries ingest reliably.
Graph quality audit
Measures coverage, orphan and duplicate nodes, and provenance gaps, and can write a coverage report after ingest.
Benefit: see the health of your knowledge graph.
Spreadsheet dual-write
Row groups become searchable document evidence, and CSV/TSV/XLSX sheets are also registered in PostgreSQL when the tables store is configured.
Benefit: the same data supports text and SQL questions.
Entity resolution
Normalizes, blocks, scores and links likely aliases with SAME_AS / canonical IDs.
Optional: enabled by feature flag.
Temporal statements
Materializes fact history as statements with supersession, conflicts and current projections, with current / as-of / history question handling.
Optional: enabled by temporal flags.
Sources & connectors
Browser upload
Stage files and launch guarded ingest jobs with progress, queue position, ETA and per-job status. Queued or running jobs can be cancelled.
Benefit: ingest without touching a terminal.
MongoDB, Google Drive & SharePoint
Connection testing, collection or cloud file discovery, staged ingest, sync and saved checkpoints. MongoDB sync is read-only snapshot comparison.
Benefit: pull knowledge from where it already lives.
Protected connector credentials
A secret store and redaction avoid returning raw connection secrets through connector views and logs. Local file sources detect create, update and delete.
Benefit: connections without leaking secrets.
Retrieval & answers
Hybrid retrieval
Qdrant dense search, Neo4j Lucene BM25 and graph Personalized PageRank, with entity linking and metadata filters.
Benefit: semantic, exact and connected evidence in one pass.
Fusion & reranking
Reciprocal Rank Fusion plus local cross-encoder sentence-window MaxP reranking, with deduplication, document diversity and neighbouring chunks.
Benefit: the most relevant passages reach the model.
Citation validation
Checks [S#] evidence references, removes invalid or unsupported cited text according to its rules, and handles refusals. A not-found gate can abstain before generation.
Benefit: answers point to real sources.
Claude or Ollama
Claude by default, or Ollama for local/offline answers, with prompt and evidence budgets and optional fallback or continuation.
Benefit: choose cloud or fully local answering.
Query-specific answers
Record list, count and aggregate paths, document summaries, library overviews, and spelling/name matching.
Benefit: the right approach for each kind of question.
Grounding verifier
Claim and citation checks are logged by default; enforce mode applies the policy to the visible answer.
Optional enforcement: set by grounding mode.
Access-aware retrieval
Tenant and visibility rules are enforced before evidence selection.
Optional: enabled by ACL flag.
Adaptive routing
A query router adjusts retrieval channel plans by question type, and an LLM router selects EASY / NORMAL / COMPLEX model choices.
Optional: enabled by router flags.
Voice response mode
Produces concise spoken-style answers and suppresses spoken citation tags.
Optional: enabled with VOICE_MODE.
Tables & analytical questions
PostgreSQL table store
Registers sheets, lists schemas and samples, and runs capped read-only SELECT queries. A separate reader connection can use a SELECT-only role.
Benefit: safe, structured analysis of spreadsheet data.
Table question routing
Detects table questions and prepares deterministic or local-model-assisted SQL for counts, filters, groups, trends and comparisons — before document RAG.
Benefit: numeric questions get computed answers.
File & table comparison
Compares versions and rows, including added, removed or changed records with evidence. Eligible follow-ups stay on the prior table or comparison.
Benefit: see exactly what changed between files.
Conversation & memory
Chat routing
A hybrid router separates smalltalk, conversation metadata, general questions, follow-ups and document questions using regex, local MiniLM intent and local Qwen.
Benefit: each message takes the right path.
Follow-up understanding
Handles repeats, reformats, subject corrections and acceptance of suggested questions using saved conversation state.
Benefit: natural multi-turn conversations.
Long-term memory
Memory nodes in Neo4j, isolated from document entities, supply bounded, non-citable context. List, create, update, archive, restore and consolidate memories.
Benefit: the assistant remembers without polluting evidence.
Operations & interfaces
Web, API & CLI
A browser app, an ingest and query CLI, and a separate query API exposing /health, /ready, /metrics and /query.
Benefit: people and automation share one engine.
Local models & Docker Compose
Local BGE embeddings and MiniLM reranking (Ollama embedding alternative). Compose defines Neo4j, Qdrant, PostgreSQL, API, UI and an optional Ollama profile.
Benefit: run the full stack on your own infrastructure.
Health, metrics & tooling
Health collection, Prometheus metrics, request tracing, optional OpenTelemetry, security audit records, plus backup, migration, retention, evaluation and benchmarking tooling.
Benefit: operate and measure it like production software.
Advantages
What customers gain in practice
Practical outcomes supported by the product — grounded in what is implemented today.
One searchable knowledge base
Documents, scans and spreadsheets land in a single graph and vector index, so questions span the whole library.
Traceable answers
Citations are validated against selected evidence, and the graph explorer shows the subgraph behind an answer.
Right tool for each question
Table questions run as read-only SQL, document questions use hybrid retrieval, and chat routing handles the rest.
Local-first options
Local embeddings, local reranking and Ollama answering support deployments that keep data on your own infrastructure.
Cost visibility
Query Economics reports measured answer tokens and cost against a full-document baseline and context reduction when comparable.
Controlled configuration
The System Config dashboard edits approved flags with validation, backups, rollback, audit and restart-pending status.
In the browser
What users get in the Answerra app
A FastAPI-backed web application with upload, Ask, Check graph, documents, connectors, tables and settings.
Streaming cited answers
For: anyone asking questions of the library
- Retrieval and answer events streamed live (SSE)
- [S#] citations linked to source evidence
- Follow-ups, reformats and suggested questions
- Chat sessions: list, rename, delete, bulk-delete
Graph explorer
For: analysts reviewing how an answer was formed
- Cytoscape view of the answer subgraph
- Evidence links back to documents and chunks
- Entities, mentions and fact edges with provenance
Table analysis
For: teams working with CSV, TSV and Excel data
- List tables with schema and sample rows
- Counts, filters, groups, trends and comparisons
- Added, removed and changed records between versions
- Sticky follow-ups on the same table
Usage, feedback & configuration
For: operators and administrators
- Query Economics: tokens, cost and context reduction
- Feedback ratings/comments and evaluation runs
- Memory management and status metrics
- System Config dashboard with rollback and audit
Answer with sources
Question routed · dense + BM25 + graph retrieval · fused and reranked · answer generated · citations validated
DenseQdrant
BM25Neo4j Lucene
GraphPageRank
[S#]Citations
Illustrative layout only — not a live answer. Real answers cite selected evidence passages, and the not-found gate can abstain when evidence is missing.
How it fits
Where Answerra sits in your stack
Answerra runs alongside your data stores and models — and is clear about what is on by default and what you enable.
Deploy
Start Neo4j, Qdrant, PostgreSQL, the API and UI with Docker Compose, with an optional Ollama profile. Scripts cover model download, installation and UI start/stop.
Ingest
Upload files in the browser, run the ingest CLI, or sync MongoDB, Google Drive and SharePoint sources. The normal ingest path is batch based.
Ask & analyze
Ask in the web UI, the query CLI or the separate query API. Table analytics need a configured PostgreSQL tables store; Ollama modes need a reachable Ollama service.
Know the scope
Temporal history, entity resolution, ACL retrieval and routing require their feature flags. The web UI and API do not provide a general user login, and cloud connector permissions are not carried into ACL metadata.
Who benefits
Built for the people who need answers from their data
Knowledge workers & analysts
Ask questions across documents and scans and get answers with citations instead of reading every file.
Data & operations teams
Query spreadsheets with counts, trends and comparisons, and see added, removed or changed records between versions.
AI & ML engineers
Tune hybrid retrieval, reranking and routing, and measure changes with the evaluation harness and model benchmarking.
Platform & DevOps teams
Run the stack with Docker Compose and operate it with health checks, Prometheus metrics, tracing, backup and retention tooling.
Administrators
Change approved settings through the System Config dashboard with validation, backups, rollback and an audit trail.
Technology leaders
Track answer token usage and cost through Query Economics, and choose between Claude and local Ollama models.
See Answerra on your own documents
Talk with AntHive Tech about a demo, a pilot, or how Answerra fits your knowledge and data workflow.
General: info@anthive.tech · Support: support@anthive.tech