How it works
How MITHUNAI works: from a source to a cited answer
How content is ingested, how a question is retrieved and answered or declined, how each organisation's data is kept apart, and the design decisions behind each step, explained in plain words for evaluators and engineers.
Last updated
What You Do vs. What MITHUNAI Does Deeply
You provide the knowledge and choose your channels. MITHUNAI handles extraction, secret redaction, embedding, the security checks and the citations.
Knowledge Ingestion
Setting up your unified company documentation and repository index.
Paste your GitHub repo URL or upload your PDFs and docs.
No manual chunking and no embedding configuration.
Multi-format extraction & pgvector HNSW indexing.
Extractors pull text from each format, secrets are redacted, GatewayEmbedder generates vectors, and pgvector builds the HNSW index.
Deployment & IDE Setup
Making answers reachable to your engineers, support staff, and visitors.
Add MCP config to Cursor or paste the widget script tag.
One JSON file in .cursor/mcp.json or one script tag on your public documentation site.
Perimeter defense & identity token minting.
Enforces request-time HTTP Origin allowlists, gives each visitor a short-lived token that can only send messages, and keeps visitor sessions apart.
Querying & Verification
Receiving answers you can verify against source code and documentation.
Ask questions naturally in plain language or code.
Query how functions work, ask policy details, or request troubleshooting steps.
Reciprocal rank fusion and a grounding check.
Fuses lexical and vector search. If the evidence holds, streams answers with citations. If missing, honestly declines rather than hallucinating.
Security & Support
Maintaining complete control, audit logs, and compliance.
Review audit logs and triage unanswered questions.
Inspect where your docs have gaps and invite support staff when needed.
Query-time tenancy & consented support sessions.
Filters every query by the caller's organisation, sends no telemetry by default, and enforces time-limited ActorType.SUPPORT sessions.
Enterprise System Architecture & Data Flow
Detailed technical schematics showing how data moves from customer sources into secure partitions, and how queries are evaluated against strict grounding invariants.
End-to-End Enterprise Architecture Topology
- • Developer IDE (MCP)
- • Web Widget (.js)
- • REST / SSE Stream
- • Admin Console
- • Origin Allowlist
- • Secret Redactor
- • Prompt Fencing
- • Bearer Token Auth
- • Model Gateway
- • Multi-Format Extraction
- • Lexical + Vector RRF
- • Grounding check
- • pgvector HNSW
- • Per-organisation isolation
- • Redis Sessions
- • No Default Telemetry
How grounding and honest abstention work
Hybrid RRF searches indexed chunks inside the caller's organisation.
assess_evidence() checks whether the passages support an answer.
Every Platform Feature, Architecturally Explained
Explore how each delivered capability in the MITHUNAI platform works, what you do as a customer, and the architectural guarantees that enforce it.
Universal Source Ingestion
What you do (customer experience)
Customer connects any of sixteen kinds of source: a GitHub repository with its issues, pull requests and discussions, a documentation site or sitemap, uploaded files, an OpenAPI schema, a public help centre or forum, or, through a connection saved once, a private repository, Confluence Cloud, Notion, Jira or Zendesk tickets (built; not yet tried against a real account).
What MITHUNAI does deeply (engine automation)
Engine extracts text from code, Markdown, HTML, PDF and DOCX, redacts secrets, generates embeddings, and indexes into pgvector.
How it is built
Connectors unified behind a single SourceConnector port. Ingestion jobs are asynchronous Celery tasks with SHA-256 content deduplication.
Key Architecture Decision Records (ADRs)
Every critical security boundary and product invariant in MITHUNAI is backed by an authoritative Architectural Decision Record.
Honest abstention is the default
Establishes that the answering engine must explicitly state "I do not know" when the indexed corpus lacks sufficient evidence, rather than extrapolating or guessing.
Widget Visitors as Non-Human Principals
Defines widget deployments as non-human principals authorized through the existing pipeline. Request-time origin allowlists stop cross-site embedding attacks.
Read-Only Model Context Protocol Tool Server
Guarantees that external IDE clients cannot execute write, delete, or mutation requests over MCP. Tools are strictly confined to reading indexed knowledge.
Consented Read-Only Support Access Sessions
Allows MITHUNAI support engineers to assist customers only with explicit opt-in consent. No credentials are shared, and every support access is logged.
How MITHUNAI Works: Complete Architecture & Workflow Guide
Understand exactly what you do as a customer, what our engine executes behind the scenes, and how every answer comes with citations you can check, or does not come at all.
Connect Your Knowledge
Provide your GitHub repository URL, public documentation website, XML sitemap, or upload PDF, Word (.docx), and text files.
Background Ingestion & Chunking
Extractors pull text from code, Markdown, HTML, PDF and DOCX, secrets are redacted first, then chunks are embedded and indexed in pgvector with HNSW.
Choose Your Integration Channel
Developers connect Cursor, VS Code or Claude Code via MCP; teams embed the web widget snippet on customer sites or call versioned REST/SSE APIs.
Perimeter Defense & Identity Binding
Verifies request-time HTTP Origin allowlists, gives each visitor a short-lived token that can only send messages, and resolves every API or MCP call to the organisation of its key.
Ask Questions in Natural Language or Code
Engineers query architecture RFCs; customer support reps ask policy questions; website visitors query feature docs.
Hybrid Retrieval & Grounding Evaluation
Executes reciprocal rank fusion (lexical + pgvector). If the evidence holds, streams answers with citations to the source passages. If missing, honestly abstains.
Audit, Scale & Request Support
Inspect unanswered question analytics, verify data boundaries, or invite MITHUNAI engineers for support sessions without sharing credentials.
Query-Time Tenancy & Consented Sessions
Filters every query by the caller’s organisation, sends no telemetry by default, and provides time-limited, consented support access.
Grounded answers from your own knowledge, a chat workspace for your team and one model gateway, on one account.
A 30-minute walkthrough on your own use case: connect a source, ask it questions, check the citations and see what it does when the answer is not there. Or ask the public playground now, with no account.