Yes. I think there is a publishable paper here, but I’d frame the motivation more carefully than “researchers know this; practitioners don’t read the papers.”

There is evidence for a real gap, but it seems to come from several mechanisms.

First, the relevant knowledge is split across fields. Interactive IR, information seeking, HCI, visual analytics, end-user programming, mixed-initiative systems, provenance, and AI have developed partly separate research traditions. Work on interactive IR has explicitly noted the limited overlap between system-oriented IR, interactive IR, and information-seeking research. (ScienceDirect)

Second, the dominant evaluation methods often favor systems that are easier to measure offline. Classical IR evaluation reduces the problem to something like query → ranking → relevance, while interactive systems involve reformulation, learning, changing goals, and different users. Interactive evaluation is richer but also substantially more expensive and harder to reproduce. (ScienceDirect) Bates identified the underlying mismatch decades ago: real information needs change as people encounter information, while classical retrieval evaluation tends to assume a stable need. (GSEIS Pages)

Third, there is an organizational boundary. A study of 21 industry practitioners across 14 organizations found that conventional separation between AI engineering and UX design breaks down for Human-AI systems: interface requirements affect model behavior, data, and system components, so teams created ad-hoc representations to communicate across those boundaries. (arXiv) This is quite close to your intuition about practitioners and organizational decision-making, but it suggests a structural problem beyond people simply not reading HCI literature.

Fourth, many of these ideas cost more to implement. A chat product needs:

input
→ model
→ output

What we’re describing needs something more like:

persistent state
+ incremental computation
+ provenance
+ representations
+ branching
+ multiple views
+ interaction design
+ latency management
+ correction semantics

That engineering-cost explanation is my inference, rather than something the literature above directly establishes.

And there is an important counterexample: when researchers have actually built systems around these principles, the results often look a lot like what impressed you about Matrix. AI Chains, for example, let people inspect and modify intermediate LLM operations; in its 20-person study, chaining improved task outcomes and increased transparency, controllability, and the users’ sense of collaboration. Users could compare alternative branches and debug individual steps. (DOI)

So I think the paper’s motivating question could be:

We already know a great deal about designing interactive systems for uncertain, evolving information work. Why do contemporary LLM-based retrieval systems so often reduce the interaction to a query box followed by an answer?

That is a much stronger opening.

Your Hebbia observation then works as the contemporary trigger. What caught your attention was specifically that Matrix exposes intermediate computation, makes it manipulable, supports multi-stage analysis, and turns the expert into an operator of the process.

I would make the first publication a position/synthesis paper, rather than claiming a novel empirical theory.

A possible title:

Beyond the Query–Answer Interface: Established Principles for Interactive AI Information Systems

Or, more specifically:

From Retrieval to Interactive Investigation: Human-Centered Principles for RAG and Query Understanding

The second is probably closer to what you actually want to say.

The core thesis could be:

Current RAG systems usually treat information access as an automated transformation from a user query to retrieved context and an answer. Decades of research in interactive information retrieval, HCI, sensemaking, mixed-initiative systems, visual analytics, end-user programming, and provenance suggest a different model: information seeking is an iterative process in which users and computational systems jointly construct and revise representations of the information need, evidence, and emerging understanding. We synthesize these traditions into a set of design principles and show how they change the architecture of RAG and query-understanding systems.

I’d structure the paper in five parts.

1. The problem: RAG inherited the wrong interaction abstraction#

Start from the dominant pattern:

query
 ↓
retrieve
 ↓
generate
 ↓
answer

Even sophisticated implementations largely make improvements inside this pipeline:

rewrite
decompose
hybrid retrieval
rerank
GraphRAG
agentic search
...

but the user commonly still experiences:

ask
 ↓
wait
 ↓
answer

The point shouldn’t be that such RAG is technically primitive. The issue is that its interaction model hides the evolving investigation.

Contrast:

AUTOMATED PIPELINE

user → query → opaque process → answer

with:

INTERACTIVE INVESTIGATION

             ┌── interpretations
             ├── questions
user ↔ state ├── retrieval operations
             ├── evidence
             ├── hypotheses
             └── conclusions
                   ↕
                 agents

2. Short literature review: the ideas already exist#

I’d keep this section deliberately concise and organized by design problem, not chronologically.

You have good anchor sources.

Shneiderman gives visible objects and rapid, incremental, reversible interaction. (UMD Computer Science)

Horvitz gives mixed initiative as a coupling of automation and direct manipulation, with each participant contributing when appropriate. (DOI)

Bates gives the evolving information need. (GSEIS Pages)

Interactive IR has explicitly studied query reformulation and interfaces for helping users alter their queries. (ScienceDirect)

Nardi and Miller give us an especially useful precedent for the computational surface: a small number of high-level, domain-specific operations combined with strong visual structure. (Miramontes Interactive)

Distributed-representation research makes an even deeper claim: the form of an external representation affects what information can be perceived and what processes and structures can be discovered. (ScienceDirect)

Interactive ML demonstrates the value of letting users inject knowledge during model construction rather than confining them to preparing input before an automated process. (ScienceDirect)

And W3C PROV gives an established formal vocabulary for tracing entities, activities, agents, uses, generations, and derivations. (W3C)

3. Synthesize the literature into principles#

I’d be careful here: the claimed contribution is not that each principle is new.

The contribution is:

a synthesis for LLM-mediated information systems, plus architectural consequences.

Perhaps seven principles:

P1 — Externalize the evolving problem state.
Important interpretations, questions, evidence, hypotheses, and decisions should exist as persistent objects.

P2 — Support mixed initiative.
Human and AI should both be able to introduce questions, revise representations, perform operations, and request clarification.

P3 — Make operations incremental, inspectable, and reversible.
The user should see the consequences of an intervention and be able to revise it locally.

P4 — Treat the information need as evolving.
Query understanding is an iterative process whose previous states and branches remain available.

P5 — Provide composable operations over typed information objects.
Search, decomposition, evidence extraction, comparison, critique, etc. should form a small computational vocabulary.

P6 — Separate shared investigation state from its views.
Matrix, graph, source view, evidence table, timeline, and conversation are coordinated projections of the same state.

P7 — Preserve provenance and epistemic status throughout computation.
Claims, transformations, evidence, uncertainty, source authority, and human/AI interventions retain their derivation.

That is already enough. I wouldn’t try to publish 15 principles.

4. Apply the principles to RAG#

This becomes the concrete technical section.

Conventional RAG:

Q
 ↓
query embedding
 ↓
top-k chunks
 ↓
LLM
 ↓
answer

Interactive investigation RAG:

                    Raw query
                        │
                        ▼
              Information-need state
             /         |          \
     interpretations   QUDs    evidence needs
             \         |          /
                        ▼
                Retrieval planning
          ┌─────────────┼─────────────┐
          │             │             │
     raw-text       question       concept/
     retrieval       retrieval      graph
          │             │             │
          └─────────────┼─────────────┘
                        ▼
                 candidate evidence
                        │
              human ↔ inspect/revise
                        │
                 evidence model
              /          |          \
          supports   contradicts   missing
              \          |          /
                        ▼
                     synthesis

And this is where our representation-relation work becomes useful.

Instead of only:

similarity⁡(E(q),E(d))

you have an inspectable retrieval plan:

(query projection,corpus projection,relation,evidence need)

For example:

QUD
 × inferred-QUD representation
 × same-question relation

question
 × hypothetical-question representation
 × answerability relation

claim
 × proposition representation
 × support/contradiction relation

concepts
 × concept graph
 × associative-path relation

That gives the paper a concrete contribution beyond generic “make AI interactive.”

5. Apply it specifically to query understanding#

This may actually be the strongest worked example.

Current:

user query
    ↓
hidden query rewriting/decomposition
    ↓
retrieval

Proposed:

user query
    ↓
┌──────────────── Information Need ──────────────────┐
│                                                   │
│ Interpretation A            Interpretation B       │
│ confidence                  confidence             │
│ goal                        goal                   │
│ QUD                         QUD                    │
│ evidence needed             evidence needed        │
│ assumptions                 assumptions            │
│                                                   │
└───────────────────────────────────────────────────┘
                    ↕
                  user
                    │
                    ▼
              retrieval probes
                    │
                    ▼
                 evidence
                    │
          ┌─────────┴─────────┐
          │ evidence changes  │
          │ understanding     │
          └─────────┬─────────┘
                    ▼
          revised information need

Conversational-search research already establishes that underspecified queries can benefit from mixed-initiative clarification. (aclanthology.org)

Our extension is:

clarification does not have to consist solely of the system asking another textual question.

The system can expose an editable interpretation:

I think you're trying to:

[x] recover an earlier thought
[ ] investigate the topic generally

Important concepts:
[ embeddings ] [ similarity ] [ relevance ]

Evidence wanted:
[x] original occurrence
[x] later refinements
[ ] external literature

The user may correct that faster than having three more conversational turns.

That is where direct manipulation, mixed initiative, interactive IR, and LLM query understanding finally connect.

And I would end the paper with an evaluation agenda, because otherwise reviewers can reasonably ask whether these principles merely produce a more complicated UI.

The hypotheses become testable:

H1  Interactive query models improve retrieval success
    on vague and exploratory information needs.

H2  Exposed intermediate representations let users
    diagnose retrieval failures more accurately.

H3  Local correction reduces recovery cost after an
    incorrect AI interpretation.

H4  Persistent investigation state improves performance
    on multi-stage information-seeking tasks.

H5  Experts benefit more from controllability as domain
    ambiguity and consequence increase.

Metrics shouldn’t just be retrieval nDCG or answer correctness. Measure:

task success
time / interaction cost
number of unrecovered failures
quality of final evidence set
ability to detect system error
correction cost
user understanding of system state
confidence calibration
provenance recovery

That also connects back to the historical problem: interactive IR evaluation has long argued that ranking metrics alone don’t capture user effectiveness and usefulness. (ScienceDirect)

So I think the publication should make one central argument:

LLMs made sophisticated information processing cheap enough to automate, but automation does not remove the need for decades-old principles of interactive information work. It makes those principles more important, because much more consequential computation is now hidden behind a conversational interface.

Then RAG/query understanding becomes the worked demonstration of what happens when we take those principles seriously.

This is a credible position paper because we wouldn’t claim that “we invented interactive AI.” The novelty would be the synthesis, translation into explicit architectural principles, and their application to modern RAG/query-understanding pipelines.