The Drag in AI-Assisted Web Research

Problem-and-discussion record. The convergent solution pattern lives in [[Source-Curation-Gate]] (blueprint). This issue is the why it hurts, the what we already have, and the open decisions. Captured per the pseudomonorepos REFLECT phase so the grilling isn’t lost.

The problem (operator’s words)

We use AI-assisted web research and AI-generated, research-driven content methods. We are likely not applying best practices. Either way:

• We get a lot of hallucinated sources.
• We get noise when there should be signal.
• Going through all the sources is time-consuming and beyond frustrating.
• The thing that should save time instead creates a ton of drag — to the point where Google Search + Google Docs manual content development might be more time-efficient.

This shows up in two of our apps that tackled it differently: augment-it (primarily via a “pack-runner” UI) and memopop-ai (primarily via the filesystem). dididecks-ai has the corpus + citation-model side. We want to converge a third, independent but very similar surface that: takes an LLM-generated list of web resources (usually Perplexity), treats each entry as an object that can be sorted, edited, deleted, or archived, can launch a SearXNG search per entry to pull metadata from the most relevant results, and lets relevant results be added back into the list.

The reframe — the drag is structural, and we already diagnosed it twice

The drag is not a prompt-quality problem. It has two structural causes, both already on record in our own corpus.

Cause 1 — retrieval and generation are conflated

One model call asked to both search and write invents URL-shaped tokens to fill citation-shaped holes — that is the smoothest completion. Diagnosed in memopop-ai/context-v/explorations/Separating-Retrieval-from-Generation-in-Agent-Pipelines.md (2026-06-08). Production evidence: 65 fabricated example.com URLs in one Panthalassa memo (v0.0.2). Operator verdict on record: “Perplexity completely hallucinates sources — cannot be trusted at all.”

The remedy is a Harvester (finds + validates URLs, emits a corpus with stable IDs) / Writer (structurally disarmed, cites by ID only) split.

Cause 2 — triage happens at the wrong unit and the wrong time

Reviewing sources per-URL, post-generation, after fabrications are already woven into prose. augment-it ran the experiment and measured the cost: the OfficialPulse audit hit a 99.7% reject rate (1,982 responses, 3 accepts), root-caused not to the volume of triage but to the unit of work being wrong (per-URL when it should be per-source-object). The drag is a units bug.

Cause 3 — Perplexity is being used against its grain

Sonar is a generative-search model: genuinely good at synthesized commentary (“what does a well-read generalist think about this deal”), catastrophic at load-bearing citations. memopop-ai/context-v/explorations/Human-Curated-Source-Sets-and-Per-Firm-RAG-for-Memo-Narrative.md (2026-05-22) already made the call: keep the Perplexity output as a quarantined “AI take” exhibit (0-ai-take.md), and no URL from it ever propagates into the cited corpus.

The convergent fix already exists three times over

All three apps independently grew the same organ: a human curation gate between retrieval and generation that produces a ranked, pruned, provenance-stamped source set, which becomes the closed corpus a structurally-disarmed writer cites from (Anthropic search_result blocks bind citations to a block index, not a URL string). The keystone theory — facts are retrieved, framing is generated — is in ai-labs/context-v/explorations/Corpus-Grounded-Generation-of-Decks-and-Memos.md (2026-06-19). The UI the operator described is the missing extraction of a pattern built three times.

The prior art (the reveal)

Keystone documents

Half-built three times

memopop (filesystem-primary):

• Curating-only-valid-Sources-across-Runs.md (2026-06-08) — the verdict ladder (verified-accessible, verified-gated, thin, title-swapped, verified-via-republish) + heuristics (soft-404, paywall, title-swap, body-length < 2KB, known-hallucination URL shapes). Pass A (dedup) / B (validity) / C (human override).
• Live data model: apps/memopop-orchestrator/io/<firm>/deals/<deal>/inputs/Sources.md — YAML frontmatter per source: url, title, publisher, sections[], rank, sensitivity, note.
• Output artifacts: redacted-hallucinations.md, source-validation-log-cleanup_sections.json.

augment-it (UI-primary — closest to the target UI that already runs):

• Search-Providers-as-First-Class-SearXNG-Default.md — SearXNG already the default connector; Tavily for content-RAG. Seam at services/social-search/src/connectors/{tavily,searxng,serpapi,gdelt,...}.ts.
• apps/pack-runner/App.svelte + ConnectorPalette.svelte — per-row, per-provider re-fire. The “launch a SearXNG search from a list entry” mechanism, already coded.
• services/social-search/src/scoring.ts — confidence pill 0–100 (tier-1 URL-shape +60, name-exact +30, recency +10, sibling-ambiguity caps 60); verification.ts — hostname-whitelist + URL-normalization dedup key.
• Response-Reviewer-Shell-and-Content-Reader-Mode.md, Enhanced-Records-List-and-Promotion-Checkpoint.md — per-object triage + record identity (record_uuid) + promote-subset + archived flag.
• Per-Record-Iteration-as-Primary-Surface-for-Pack-Fires.md — validate on 4–8 records by hand before bulk fan-out.

dididecks (corpus-primary):

• Substantiation-Corpus-Data-Model.md (2026-06-07) — CorpusItem (pointer-only, provenance, superseded_by).
• Dididecks-AI-DD-Ready-Citation-and-Source-Access.md — receiving-side verifiability; archive-on-cite for link-rot.

Shared substrate (ai-labs parent):

• Chat-As-Verb-Surface-Patterns.md / Per-App-Workspace-Conventions.md — CapabilityResult<T> envelope + entity.verb registry the gate’s source.* verbs register against.

The packaging discipline — knots-style, NOT a shared dependency

Hard constraint from the operator (2026-06-27): dididecks-ai, augment-it, and memopop-ai are fundamentally different apps and stay that way for now. ai-labs is a pseudomonorepo. The cross-app sharing model is the astro-knots knots lesson: @knots/* as imported dependencies failed (too much abstraction overhead; UI components especially, because “sites diverge too much in design for generic components to be useful”). What survived: co-located development + shared context-v/ documents + selective copy-from sample code, cross-referenced with @context-v/*. Only genuine dependencies get published (@lossless-group/lfm is the one real package).

Therefore the Source Curation Gate is shared as a pattern, not a package:

• The blueprint ([[Source-Curation-Gate]]) + any sample code live at the ai-labs parent level.
• Each app re-implements the gate against its own stack and data model, conversationally linked back to the blueprint via @context-v/* and [[wikilinks]].
• Do not propose a shared microfrontend, a published @gate/* package, or any true import dependency across the three apps. That is the move that already didn’t work.
• Convergence is achieved by shared vocabulary and shared schema shape (the Source object, the verdict ladder, the source.* verbs), not shared runtime code.

The grilling — open forks (answers fold into [[Source-Curation-Gate]])

Recommendations attached; each is meant to be argued with.

On research methodology

• G1 — The gate’s primary verb may need to be “search better,” not “triage.” Candidate-pool quality is set upstream by query construction; a gate that only polishes a bad list is lipstick. memopop admits the gate “does NOT solve the harvester’s query construction.” Recommendation: rejected sources teach the next SearXNG query (negative keywords, domain exclusions) — a query-refinement loop, not just a triage loop. Fork: is the gate a search loop or a triage tool?
• G2 — Perplexity is the worst possible seed and the task enshrines it. Its URLs are structurally fabricated. Recommendation: treat the Perplexity list as a topic-and-entity / claim extractor, discard its URLs wholesale, and use SearXNG/Valyu to find real sources for those claims. Fork: keep its URLs, or mine it for claims and throw the URLs away?
• G3 — SearXNG returns thin SERP metadata; grounding needs the fetched body. Recommendation: two-tier — SearXNG for discovery + ranking, then Jina Reader on promoted-only URLs for the verbatim body. Fork: two network tiers, or one-call retrieve+extract (Firecrawl) and eat the cost?

On RAG discipline

• G4 — “Relevant results can be added” — relevant to what, measured how? If relevance = eyeballs, it doesn’t scale; if = a model score, it drifts. Recommendation: relevance = embedding match against the specific unfilled fact-slot the deliverable needs, not against the topic. Fork: embedding-match relevance now, or human judgment at v1?
• G5 — Verbatim-or-bust at ingest. Keystone doc forbids LLM-summarizing at ingest (failure mode #3, ingest distortion; MemPalace 92.9% vs Mem0 30–45% rides on this). Recommendation: the gate may display an LLM summary as convenience, but the stored, citable artifact is always the verbatim fetched body. Fork: store verbatim only, or also store extracted “key facts” (and accept distortion)?
• G6 — Auto-validate before the human sees the list, or the drag never dies. memopop principle: “validate before the human ranks; the analyst never clicks past dead URLs.” Recommendation: the verdict ladder runs automatically on import; the human spends attention only on rank + relevance, never on liveness. Fork: auto-validate as a hard precondition, or analyst-triggered at v1?

On scope and architecture

• G7 — Does the gate end at “promoted corpus,” or own generation too? Recommendation: stop at the promoted, grounded corpus (Sources.md + optional Chroma collection); generation is a separate consumer — that is what keeps it droppable into all three apps. Fork: stop at corpus, own generation, or also own upstream query refinement (G1)?
• G8 — “Independent” packaging. Settled by the knots constraint above: shared pattern + sample code at the parent, re-implemented per app, never a shared dependency. The remaining sub-fork is only where the first reference implementation gets built (which app hosts the sample others copy from).

The four load-bearing forks to settle first

1. First reference implementation — which app hosts the sample code the others copy from? (augment-it has the most already-built: pack-runner + connectors + scoring. memopop has the data model + the live pain. dididecks has the corpus/citation model.)
2. Scope boundary (G7) — stop at corpus / own generation / own query refinement.
3. Perplexity seed handling (G2) — keep URLs (validate hard) / mine claims and discard URLs / both (sealed AI-take + mined claims).
4. Validation ownership (G6) — auto-on-import / analyst-triggered / hybrid (cheap auto, deep on-promote).

Anti-patterns (carried into the blueprint)

• Letting AI-take URLs leak into the corpus.
• Triaging post-generation, per-URL (OfficialPulse 99.7%).
• Treating rank as cosmetic (it drives hedge calibration + contradiction tie-break).
• HTTP-200 as validity (soft-404s, paywalls, title-swaps).
• Proposing a shared dependency/microfrontend across the three apps — knots already proved that fails; share the pattern, not the package.

Related

• [[Source-Curation-Gate]] — the convergent solution pattern (blueprint).
• [[Corpus-Grounded-Generation-of-Decks-and-Memos]] — the generation theory the gate feeds.
• [[Human-Curated-Source-Sets-and-Per-Firm-RAG-for-Memo-Narrative]] — the gate spec + provider landscape (memopop).
• [[Separating-Retrieval-from-Generation-in-Agent-Pipelines]] — harvester/writer split (memopop).
• [[Curating-only-valid-Sources-across-Runs]] — verdict ladder + heuristics (memopop).
• astro-knots/README.md — the knots lesson on shared-pattern-vs-shared-dependency.
• augment-it: apps/pack-runner/, services/social-search/src/{scoring,verification,connectors}.ts.
• memopop: apps/memopop-orchestrator/io/<firm>/deals/<deal>/inputs/Sources.md.