Future Philosophy · IllTellYou

From retrieval machine to living knowledge.

A new paradigm where the knowledge creator — not the seeker — becomes the primary subject of the system.

Interactive Context Management · Internal Vision Document


The subject changes.
Everything follows.

The entire RAG literature — including Karpathy's LLM-wiki — is designed around one subject: the knowledge seeker. The pipeline exists to serve queries. Even sophisticated wiki patterns are fundamentally about making retrieval better for the person asking.

What IllTellYou's next phase proposes is a different subject entirely: the knowledge creator as protagonist. The system's job is not to serve queries — it is to help the provider improve, maintain, and trust their own knowledge. Retrieval becomes a side effect, not the goal.

Every system today

The knowledge seeker
is the subject

  • Pipeline optimised for query quality
  • Knowledge is static after upload
  • Re-derives answers from scratch on every query
  • No accumulation, no learning, no compounding
  • Provider uploads and disappears
  • Contradictions are invisible
  • Interactions vanish after the session ends
IllTellYou · Next

The knowledge creator
is the subject

  • System supports authorship, not just retrieval
  • Knowledge is a living workspace the provider inhabits
  • Knowledge compiles once, compounds over time
  • Provider can test, interrogate, and correct their own content
  • Student interactions feed back into improvement signals
  • Contradictions are surfaced and flagged automatically
  • Sessions become a memory the provider can mine

Three feedback loops
that don't exist in RAG.

Standard RAG has one relationship: query in, answer out. Interactive Context Management introduces three new relationships — none of which require a student to trigger them.

🪞
Provider ↔ Knowledge

The author tests their own material

The provider interrogates their knowledge base as a student would — not to get an answer, but to see how their content is represented. Gaps, stale claims, and weak explanations become visible. Inline amendment follows naturally.

🔁
Interactions ↔ Knowledge

Student sessions improve the source

Every question asked by a student is a signal. Repeated questions reveal gaps. Poor answers reveal weak content. The system clusters these signals and presents them to the provider as prioritised improvement suggestions — without the provider needing to monitor sessions directly.

Knowledge ↔ Knowledge

The system notices what the author misses

When a new source contradicts an existing one, or when a concept is mentioned across multiple sources with inconsistent definitions, the system flags it. No external audit required. The knowledge base acts as its own editor.


Four primitives of
knowledge authorship.

Starting from the knowledge creator's needs — not from retrieval requirements — yields four atomic operations. These are not features; they are the verbs of a new interface paradigm.

01 · reflect()

Reflect

"Show me what you know about X."

The provider asks the system a question not to get an answer for a student, but to see how their knowledge is currently represented. The output is a mirror for the author — revealing how the material would respond to a real query. The starting point for any amendment session.

# What does my knowledge currently say? def reflect(teacher_key: str, topic: str) → ReflectionResult
02 · challenge()

Challenge

"What is inconsistent, missing, or contradictory?"

The system proactively analyses the knowledge store and surfaces tensions — without being asked. No query needed; the system acts as an editor that never gets bored. Particularly valuable for wiki and multi-source memo categories where contradictions naturally accumulate over time.

# Contradictions, gaps, stale claims def challenge(teacher_key: str) → list[ChallengeItem]
03 · learn_from_interactions()

Learn

"What did my students ask that I didn't answer well?"

The system processes session history, clusters questions by topic, identifies questions that received weak or incomplete answers, and presents them as ranked improvement suggestions. The provider decides whether to act. This transforms student sessions from ephemeral conversations into a permanent improvement signal.

# What did interactions reveal? def learn_from_interactions( teacher_key: str, since: datetime) → list[LearningSignal]
04 · amend()

Amend

"Correct this — here, now, without re-uploading."

A targeted, conversational correction that the system integrates into the existing knowledge base. Not a delete-and-replace workflow. Not a full re-upload. A surgical edit triggered inline from any of the other three operations — the natural conclusion of the reflect/challenge/learn cycle.

# Apply a targeted correction def amend(teacher_key: str, amendment: Amendment) → AmendmentResult

Not every category needs
all four operations.

The primitives map onto IllTellYou's category system unevenly — and that's correct. A memo is almost entirely Reflect and Amend. A university course generates the richest signal for Learn. The architecture supports all four everywhere; which are surfaced in the UI is a category-level configuration.

Category Reflect Challenge Learn from interactions Amend
Education H M H
Wiki H M H
Memo M L
Coaching M L H M
Family M L L H

★ = Very High  ·  H = High  ·  M = Medium  ·  L = Low


ICM as a service layer
above the existing pipeline.

Interactive Context Management does not replace the RAG pipeline — it uses it as its substrate. The Has-A composition pattern already established in IllTellYou's codebase means ICM slots in cleanly without architectural rewrites.

New layer — ICMService
ICMService reflect · challenge · learn · amend
Has-A composition — existing services
QdrantContentManagementService reflect · challenge
SessionMemoryService learn from interactions
AmendmentService amend
LLMQueryService analysis calls
Existing infrastructure — unchanged
Qdrant vector store
MySQL amendments · wiki pages
Redis session memory
ElevenLabs · OpenAI voice + LLM
What's new
New serviceWikiCompilerService — background LLM calls that build compiled wiki pages from raw chunks
New serviceAmendmentService — MySQL-backed, chunk-level correction overlay injected at prompt time
ExtendedSessionMemoryService — Redis short-term already live; MySQL long-term adds GOLD persistence + clustering
New callsAnalysis LLM calls — background/on-demand, not query-time; rate-limited to teacher portal only

A third mode in the
teacher portal.

Today the teacher portal has two modes: add content, and share. ICM introduces a third — "Tend your knowledge." Three entry points, each mapping to one or two primitives, each optional without affecting the core product.

01

Test your knowledge

The provider asks a question and sees what their own knowledge base would say — not a student's query, but a provider's audit. Gaps and weak answers are immediately visible. Inline amendment available after any response.

reflect() · amend()
02

Audit for conflicts

The system analyses the full knowledge base and returns a structured report: contradictory claims between sources, outdated facts, concepts mentioned with inconsistent definitions, orphaned material with no cross-references.

challenge()
03

What did students ask?

Clustered questions from all student sessions, ranked by frequency and answer quality. The questions that revealed gaps in the material are surfaced first. Each cluster links directly to the relevant content — or to the absence of it.

learn_from_interactions()

RAG is Pro.
Living knowledge is Gold.

The raw Qdrant RAG pipeline — already built — defines the Pro experience. ICM defines Gold: knowledge that improves itself over time, powered by background LLM analysis and persistent memory.

Free

Static knowledge

  • Upload, vectorise, share
  • 20 queries / day
  • No ICM primitives
  • Session memory not persisted
Pro

Full RAG pipeline

  • Hybrid search · voice cloning
  • Unlimited content sources
  • Reflect + Amend primitives
  • Current-session memory
Gold

Knowledge that compounds

  • All Pro features
  • Persistent long-term memory
  • All four ICM primitives
  • Challenge: auto contradiction detection
  • Learn: student interaction clustering
  • Wiki compiler — compiled synthesis layer

Not a better retrieval tool.
A different kind of product.

Every competitor — ChatPDF, NotebookLM, custom RAG deployments — optimises the query path. None of them treat the knowledge creator as the primary subject. That is the opening.

"The shift from 'RAG platform' to 'Interactive Context Management' changes what we are building. Every competitor is a retrieval system. ICM is an authorship system. The students benefit downstream — but the provider is the one the platform serves most deeply."

— IllTellYou · Internal architecture notes, 2026

Every RAG platform

  • Knowledge is a static artefact
  • Provider uploads and disappears
  • System optimised for query quality
  • Interactions vanish after sessions end
  • Contradictions are never surfaced
  • No feedback loop to the creator
vs

IllTellYou · ICM

  • Knowledge is a living workspace
  • Provider inhabits and tends the system
  • System optimised for authorship quality
  • Interactions become a permanent improvement signal
  • Contradictions are automatically flagged
  • Every session makes the knowledge base stronger
🔒

Internal vision document

This page describes the philosophy behind IllTellYou's future roadmap. It is not public-facing. If you're reading it, you're already part of the story.