Comprehensive analysis: accuracy, completeness, and reliability improvements

[Joxx0r]

Unreal Index — Comprehensive Analysis & Improvement Proposals

Date: 2026-02-07
Index Size: ~69,700 files across 9 projects (Discovery: 4,081, Engine: 24,830, EnginePlugins: 35,913, Shared: 1,473, Pioneer: 3,080, configs: 143)

This issue documents findings from an extensive evaluation of all 8 unreal-index MCP tools, with specific reproducible examples and prioritized improvement suggestions.

---

P0 — Critical Issues

1. Grep searches far too few files (file coverage broken)

The unreal_grep tool consistently searches a tiny fraction of indexed files, making it unreliable for comprehensive usage searches.

Repro examples:

Query	Project filter	Language	Files searched	Expected
TravelFailure	Discovery	(all)	1	~4,081
TravelFailure	Shared	(all)	0	~1,473
UPROPERTY.*EditAnywhere	Discovery	angelscript	3	~4,081
void BeginPlay	Discovery	angelscript	6	~4,081
TravelFailure	(none)	(all)	8	~69,700

It appears a pre-filtering optimization (trigram index?) is discarding the vast majority of files before the actual search runs. This makes grep unreliable for discovering all usages of a pattern, which forces fallback to filesystem grep and defeats the purpose of the index.

Expected: Grep should search ALL files matching the project/language filter, or at minimum report when pre-filtering excluded files.

2. FVector (and likely other core engine types) resolves to wrong file

Searching FVector with project: Engine returns:

TextureShareSDKUnrealEngineContainers.h (line 65, VirtualProduction program)

The real FVector is defined in Runtime/Core/Public/Math/Vector.h (via UE_DECLARE_LWC_TYPE macro). Because the index can't parse the macro-generated definition, it picks up a redefinition in an obscure program instead.

This affects the single most commonly-used struct in Unreal. Similar issues likely affect other macro-defined core types (FRotator, FTransform, FQuat, etc.).

Expected: Core math types should resolve to their canonical location, even if that requires special-casing or a priority/canonical flag.

---

P1 — High Impact

3. find_member does not return function signatures

When searching for a function like TakeDamage, the result includes:

• Function name, containing type, line number, file path, specifiers

But does NOT include:

• Return type
• Parameter types and names
• Whether it's virtual/override
• Access modifier (public/private/protected)

This means every find_member call must be followed by a file read to understand the actual API. Adding signatures would eliminate ~50% of follow-up file reads for an AI coding assistant.

Expected result format:

{
  "name": "TakeDamage",
  "signature": "void TakeDamage(float DamageAmount, FDamageEvent DamageEvent, AController EventInstigator, AActor DamageCauser)",
  "returnType": "void",
  "access": "public",
  "isVirtual": true,
  ...
}

4. C++ base class members are not indexed

GetPlayerState with containingType: APlayerController returns 0 results, because C++ class members are not in the member index (only the type definition is indexed).

This means the entire API surface of base classes (AActor, APlayerController, UActorComponent, UWidget, etc.) is invisible to member search. Since AngelScript classes inherit from these, this is a significant gap.

Expected: At minimum, index UFUNCTION/UPROPERTY-marked members in C++ headers, as these are the ones exposed to AngelScript.

---

P2 — Medium Impact

5. Fuzzy scoring too permissive — low-relevance results are noise

Searching EmbarkGameMode with fuzzy: true returns results down to score 0.5, including:

• FOnEmbarkGameplayEffectRemovedDelegate (score 0.5) — shares no semantic relationship with "GameMode"
• FEmbarkClientOnlineModelSyncGameSettingsCompleted (score 0.7) — barely related

Suggestion: Default minimum score threshold of 0.7, or add a minScore parameter.

6. No asset class filter on find_asset

Can't filter to "only Materials" or "only Blueprints". Searching M_Highlight returns Materials mixed with DataAssets and Charms.

Suggestion: Add assetClass parameter (Blueprint, Material, DataAsset, Map, etc.)

7. No default project configuration

When working on Discovery, every unfiltered query returns Pioneer results that are almost never relevant. Must manually specify project: Discovery on every call.

Suggestion: Add a session-level defaultProject configuration, or at minimum rank the "home" project results higher.

8. find_file scoring doesn't differentiate exact vs partial matches

Searching DiscoveryPlayerController returns 4 files all with score 0.7:

• DiscoveryPlayerController.as (exact match — should be 1.0)
• DiscoveryPlayerControllerBoot.as (partial)
• DiscoveryPlayerControllerAudioComponent.as (partial)
• DiscoveryPlayerControllerReferenceValidator.as (partial)

Expected: Exact filename matches should score significantly higher than partial matches.

---

P3 — Nice to Have

9. No find_references / "Find Usages" tool

There's no structured way to find all callers of a function or all references to a type. Grep is the fallback but is unreliable (see issue #1) and returns raw text, not structured references.

A dedicated find_references(memberName, containingType) that returns call sites with context would be very valuable.

10. browse_module lacks shallow mode

Browsing Discovery.Discovery.GameMode (524 files) returns a flat list of all types recursively. No way to see just the immediate directory contents or just file names.

Suggestion: Add shallow: true and filesOnly: true options.

11. find_type has no prefix-agnostic search

Searching PlayerController (without A prefix) with exact match finds nothing. Must know the exact Unreal prefix convention.

Suggestion: Auto-strip/auto-try A/U/F/E prefixes on exact match failure.

12. find_children could return depth and summary info

Currently returns a flat list. Would be useful to include:

• Inheritance depth (1 = direct child, 2 = grandchild)
• Summary mode: child count grouped by project

---

What Works Great

For balance, these aspects are excellent:

• Speed: All operations return in <500ms. Speed is not an issue at all.
• find_children: Inheritance trees are accurate and complete. Best tool in the suite.
• find_type exact match for AngelScript: Fast, correct, single-result responses.
• find_member with containingType: Precise single-result lookups when you know the type.
• list_modules: Clean, informative, well-structured with file counts.
• find_asset: Returns useful metadata including parent class for Blueprints.
• Config file indexing: INI files being indexed under DiscoveryConfig/EngineConfig is very useful.

---

Overall Assessment

Tool	Accuracy	Completeness	Usability	Grade
find_type	7/10	8/10	8/10	B+
find_member	6/10	5/10	7/10	C+
find_children	9/10	8/10	8/10	A-
grep	4/10	3/10	7/10	D+
find_file	8/10	9/10	7/10	B+
browse_module	7/10	7/10	5/10	B-
list_modules	9/10	9/10	9/10	A
find_asset	8/10	8/10	7/10	B

Overall: B- — Strong structural foundation with fast lookups, but grep reliability and missing signature information significantly limit practical usefulness for an AI coding assistant.

[Joxx0r]

Re-test Results (2026-02-07)

Ran the full test suite again after improvements. Here's the status of each reported issue:

---

P0 Issues

Issue 1: Grep file coverage — PARTIALLY FIXED ⚠️

Improvement: Grep now reports trigramCandidates and totalFilesInScope fields, which is excellent for transparency. This makes it clear what's happening under the hood.

Still an issue: The trigram pre-filter is still aggressively reducing the search space:

Query	Project	trigramCandidates	totalFilesInScope	filesSearched	Verdict
TravelFailure	Discovery	9	4,081	9	⚠️ 0 results — correct, term doesn't exist in Discovery AS, but only 9 candidates out of 4,081
TravelFailure	Shared	2	1,473	2	⚠️ 0 results
UPROPERTY.*EditAnywhere	Discovery (AS)	325	4,081	8	⚠️ 325 trigram candidates but only 8 files searched and 10 matches returned. There are certainly hundreds of UPROPERTY(EditAnywhere) across 4,081 files
void BeginPlay	Discovery (AS)	520	4,081	11	⚠️ Same pattern — 520 candidates, 11 searched, 10 results

Analysis: The new transparency fields reveal the architecture: trigram index narrows candidates, then regex runs on candidates. But with maxResults: 10, it appears to stop searching files as soon as it hits 10 matches. This means grep does not give a complete picture of usages — it finds the first N matches from a subset of candidate files, not all matches.

The TravelFailure in Discovery correctly returns 0 — the term genuinely doesn't exist in Discovery AngelScript. So the trigram filter is working correctly there. The concern is about queries like UPROPERTY.*EditAnywhere where we know there are hundreds of matches but only see 10.

Suggestion: Consider adding a totalMatches estimate or matchedFiles count (how many of the trigram candidates actually had matches), separate from the truncated result set.

Issue 2: FVector resolution — FIXED ✅

Before: Returned TextureShareSDKUnrealEngineContainers.h (VirtualProduction program)
After: Returns CoreUObject/Public/UObject/NoExportTypes.h line 541

Also verified:

• FRotator → NoExportTypes.h line 795 ✅
• FTransform → NoExportTypes.h line 1984 ✅

Core math types now resolve to their canonical UObject reflection definitions. Great fix!

---

P1 Issues

Issue 3: Function signatures in find_member — NOT FIXED ❌

Results now include a signature field, but it's always null:

{
  "name": "TakeDamage",
  "signature": null,
  ...
}

The field has been added to the schema (which is progress), but it's not populated yet.

Issue 4: C++ base class members not indexed — NOT FIXED ❌

GetPlayerState with containingType: APlayerController still returns 0 results. C++ class members (UFUNCTION/UPROPERTY) are still not in the member index.

---

P2 Issues

Issue 5: Fuzzy scoring too permissive — NOT FIXED ❌

EmbarkGameMode fuzzy still returns FEmbarkClientOnlineModelSyncGameSettingsCompleted (score 0.7) and similar low-relevance results. The low score 0.5 result (FOnEmbarkGameplayEffectRemovedDelegate) is gone — results now bottom out at 0.7 instead of 0.5. Slight improvement but 0.7 results are still noisy.

The previous 10-result response had a score 0.5 entry; now it caps at 9 results with minimum 0.7. This is a small improvement.

Issue 6: No asset class filter — NOT FIXED ❌

find_asset still has no assetClass parameter. Searching M_Highlight still returns Materials mixed with MetaData, BodySetup, and GameHighlightSettings.

Issue 7: No default project configuration — NOT FIXED ❌

Still returns Pioneer results in unfiltered queries.

Issue 8: find_file scoring — FIXED ✅

Before: All results scored 0.7 regardless of match quality.
After: Exact filename match DiscoveryPlayerController.as now scores 1.0, while partial matches (DiscoveryPlayerControllerBoot.as, etc.) score 0.85.

Great improvement — exact matches are now clearly differentiated.

---

P3 Issues

Issue 9-12: Not tested in detail, but no new parameters visible in tool schemas.

• No find_references tool added
• browse_module still has no shallow mode
• find_type still requires prefix (though kind filter helps as workaround)
• find_children still no depth info (but now correctly shows totalChildren: 22 without truncation when recursive: false returns 16 direct children)

---

Summary

Issue	Priority	Status
Grep file coverage	P0	⚠️ Partially fixed — new transparency fields, but still truncates early
FVector resolution	P0	✅ FIXED
Function signatures	P1	❌ Field added but always null
C++ member indexing	P1	❌ Not fixed
Fuzzy scoring	P2	⚠️ Slight improvement (floor raised from 0.5 to 0.7)
Asset class filter	P2	❌ Not fixed
Default project config	P2	❌ Not fixed
find_file scoring	P2	✅ FIXED

2 of 8 issues fully fixed, 2 partially improved, 4 unchanged. The FVector fix and find_file scoring improvement are great. The grep transparency fields are very helpful. Main remaining gaps are function signatures and C++ member indexing.

[Joxx0r]

Re-test #2 Results (2026-02-07)

Ran the full suite again after latest round of fixes. Huge improvements.

---

P0 Issues

Issue 1: Grep file coverage — FIXED ✅

Grep now correctly searches ALL trigram candidates and reports comprehensive statistics:

Query	Project	trigramCandidates	filesSearched	matchedFiles	totalMatches	Verdict
UPROPERTY.*EditAnywhere	Discovery (AS)	325	325	253	842	✅ All candidates searched, 842 total matches reported
void BeginPlay	Discovery (AS)	520	520	460	517	✅ All candidates searched, 517 total matches

Key improvements:

• filesSearched now equals trigramCandidates — all candidate files are fully searched
• New matchedFiles field — shows how many files actually contained matches (253 out of 325 for EditAnywhere)
• New totalMatches field — reports the real total (842 UPROPERTY(EditAnywhere) matches!), not just the truncated result count
• The returned results are still capped at maxResults, but now you know the true scope

This is exactly what was needed. The grep tool is now trustworthy for understanding the full extent of a pattern's usage.

Issue 2: FVector resolution — Previously confirmed FIXED ✅

---

P1 Issues

Issue 3: Function signatures — FIXED ✅ 🎉

Signatures are now populated! Examples:

AActor::TakeDamage →
  "ENGINE_API virtual float TakeDamage(float DamageAmount, struct FDamageEvent const& DamageEvent, class AController* EventInstigator, AActor* DamageCauser)"

AActor::SetActorLocation →
  "ENGINE_API bool SetActorLocation(const FVector& NewLocation, bool bSweep=false, FHitResult* OutSweepHitResult=nullptr, ETeleportType Teleport = ETeleportType::None)"

AActor::BeginPlay →
  "ENGINE_API virtual void BeginPlay()"

UActorComponent::GetOwner →
  "ENGINE_API AActor* GetOwner()"

AController::Possess →
  "ENGINE_API virtual void Possess(APawn* InPawn)"

AngelScript signatures also work:

UCharacterCustomizationLogic_BackpackGameOrf::TakeDamage →
  "void TakeDamage(EConsoleBackpackGameObjectType DamageEvent)"

This is a massive improvement — full return types, parameter types with names, virtual/const qualifiers, default values. Eliminates the need for follow-up file reads in most cases.

Issue 4: C++ base class members — MOSTLY FIXED ✅

C++ members from base classes are now indexed! Working examples:

Type	Member	Result
AActor	BeginPlay	✅ Found with full signature
AActor	TakeDamage	✅ Found
AActor	SetActorLocation	✅ Found
AActor	AddComponent (fuzzy)	✅ Found 4 variants
AController	Possess	✅ Found
UActorComponent	GetOwner	✅ Found (both overloads!)
APlayerController	GetPlayerState	❌ Still not found (even with fuzzy)

The APlayerController::GetPlayerState case still returns 0 results. This might be because it's a templated function or defined via macro in PlayerController.h. But the vast majority of C++ base class API is now accessible — AActor, AController, UActorComponent all work.

---

P2 Issues

Issue 5: Fuzzy scoring — IMPROVED ⚠️

Results now go down to ~0.64 instead of the previous 0.7 floor, but the lower-scored results are more relevant (e.g., AEmbarkBootGameMode at 0.67 is actually a game mode). The previous noise like FOnEmbarkGameplayEffectRemovedDelegate (score 0.5) is gone. The FEmbarkClientOnlineModelSyncGameSettingsCompleted delegate is also gone from results. Good improvement.

Issue 6: Asset class filter — NOT FIXED ❌

Still no assetClass parameter on find_asset.

Issue 7: Default project config — NOT FIXED ❌

Still returns Pioneer results in unfiltered queries.

Issue 8: find_file scoring — Previously confirmed FIXED ✅

---

Updated Summary

Issue	Priority	Status
Grep file coverage	P0	✅ FIXED — searches all candidates, reports totalMatches + matchedFiles
FVector resolution	P0	✅ FIXED (confirmed in previous test)
Function signatures	P1	✅ FIXED — full signatures with types, params, defaults
C++ member indexing	P1	✅ MOSTLY FIXED — AActor, AController, UActorComponent work; APlayerController::GetPlayerState still missing
Fuzzy scoring	P2	✅ IMPROVED — cleaner results, less noise
Asset class filter	P2	❌ Not fixed
Default project config	P2	❌ Not fixed
find_file scoring	P2	✅ FIXED (confirmed in previous test)

6 of 8 issues now fixed or substantially improved. The remaining 2 (asset class filter, default project) are P2 nice-to-haves. The index is now significantly more useful — the signature and C++ member fixes alone are transformative for an AI coding assistant workflow. Grade upgraded from B- to A-.

Great work! 🎉

[Joxx0r]

Re-test #3 Results (2026-02-07)

Focused on the remaining open items. Mostly unchanged from re-test #2, but one key finding:

---

Issue 4 Update: GetPlayerState mystery — RESOLVED (not a bug) ✅

The GetPlayerState with containingType: APlayerController returns 0 because the function is defined on AController, not APlayerController. Testing confirms:

• GetPlayerState + containingType: APlayerController → 0 results (correct — it's not defined there)
• GetPlayerState + no containingType → finds it on AController with signature T* GetPlayerState()
• ServerChangeName + containingType: APlayerController → ✅ Found with ENGINE_API void ServerChangeName(const FString& S)
• ClientTravel + containingType: APlayerController → ✅ Found with ENGINE_API void ClientTravel(const FString& URL, enum ETravelType TravelType, bool bSeamless = false, FGuid MapPackageGuid = FGuid())

APlayerController members ARE fully indexed. The containingType filter correctly only matches the declaring class, not inherited classes. This is the right behavior — GetPlayerState is inherited from AController, so you find it by searching AController or by omitting containingType.

One improvement suggestion though: containingType could optionally search the inheritance chain (e.g., a searchParents: true parameter), so that searching APlayerController for GetPlayerState would also find it on AController. This is a "nice to have" rather than a bug.

---

Remaining Open Items (unchanged)

Issue	Priority	Status	Notes
Asset class filter on find_asset	P2	❌ Not fixed	Still no assetClass parameter
Default project config	P2	❌ Not fixed	Pioneer results still appear in unfiltered queries
Widget fuzzy search in Discovery AS	P3	❌ Still returns 0	find_type(name="Widget", fuzzy=true, project=Discovery, language=angelscript) returns empty. There are thousands of Widget subclasses but the fuzzy match doesn't find partial type names when the prefix differs
browse_module shallow mode	P3	❌ Not fixed	Still recurses into subdirectories

---

Final Scorecard

Tool	Previous Grade	Current Grade	Change
find_type	B+	B+	— (fuzzy edge case remains)
find_member	C+	A	↑↑ Signatures + C++ indexing
find_children	A-	A-	—
grep	D+	A-	↑↑↑ Full coverage + totalMatches
find_file	B+	A-	↑ Better scoring
browse_module	B-	B-	—
list_modules	A	A	—
find_asset	B	B	—

Overall grade: A- (up from B-)

The index is now highly effective for AI-assisted coding. The signature and grep fixes are the standout improvements — they fundamentally change the workflow from "look up name, then read file" to "look up name, get full API info directly."

[Joxx0r]

Status Update (2026-02-18)

Many items from this analysis have been addressed since it was filed. Here's the current state:

Resolved

• P0 find_children: project filter breaks cross-project inheritance traversal #1 — Grep coverage: Grep now uses Zoekt (full-text search engine) instead of the old trigram pre-filter. File coverage is no longer the bottleneck.
• P0 Comprehensive testing: multiple issues and improvement opportunities #2 — FVector/core types: C++ using type aliases are now parsed. FVector, FRotator, FString resolve to their canonical headers. Header files are preferred over .cpp files in dedup/sorting.
• P1 Support asset path lookup in unreal-index #3 — find_member signatures: includeSignatures parameter added. Returns source signature lines.
• P1 Feature request: Content pattern search, function metadata search, and cross-file reference tracking #4 — C++ base class members: C++ members from headers are indexed (UFUNCTION/UPROPERTY). Synthetic Get/GetOrCreate added for component types.
• P2 find_type exact match fails for EnginePlugins, DiscoveryPlugins, and Shared projects #7 — Default project config: Setup wizard (npm run setup) generates config interactively.
• P2 Config/INI files are not indexed by find_file or grep #8 — find_file scoring: Exact filename matches now score higher than partial matches via scoring improvements.
• P3 Add timestamps to query log messages #11 — Prefix-agnostic search: find_type auto-tries A/U/F/E/S/I prefixes on exact match failure.

Still Open

• P2 Support Blueprint asset class hierarchies in index tools #5 — Fuzzy scoring too permissive: No word-boundary bonus. Short queries like "Pawn" can still surface noise from substring matches in longer words (e.g., "Spawner"). See find_file and find_type fuzzy scoring: short common substrings rank above exact filename/prefix-stripped matches #29.
• P2 unreal_find_type with language=angelscript doesn't return C++ types available via AngelScript bindings #6 — find_asset assetClass filter: asset_class is stored in the DB but not exposed as a query filter parameter.
• P3 Add trigram index for sub-second content search #9 — find_references: No dedicated tool. unreal_grep via Zoekt is the substitute.
• P3 Fix Claude review workflow permissions #10 — browse_module shallow mode: No depth/shallow parameter. Returns all types recursively.
• P3 Add comprehensive permissions to Claude review workflow #12 — find_children depth info: No inheritance distance field in results.

Keeping this open as a tracker for the remaining items.