The .gitignore had three competing rules — .env.* (line 70), *.env.*
(line 259), and a stale negation !.env.example (line 71) that lost to
the later rules. Adding the negation at the end of the file so it wins.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- server_header=False — uvicorn was adding "server: uvicorn" after our
middleware ran, defeating the banner-strip in security_headers().
- proxy_headers + forwarded_allow_ips=* — so request.client.host
reflects the real client IP from X-Forwarded-For (Traefik), which
is what the per-IP rate limiter keys on.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two issues blocked the security-fix deploy:
1. Starlette's MutableHeaders has no .pop() — del-via-membership is the
correct mapping op. The previous code crashed every request with
AttributeError, taking the container's healthcheck down with it.
2. Coolify's in-container healthcheck falls back to wget which wasn't
installed in our slim base. Add wget alongside curl.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the legacy English MCP-only README with a comprehensive Hebrew
README covering the full Web + MCP product: features, architecture,
3-tier appraisals, smart "same street" inference, RTL hardening guide,
deployment topology, all REST endpoints with curl examples.
Also update pyproject.toml description to match.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
UI:
- Center-align all numeric table columns (deals + appraisals) for better
visual balance — text-start was crowding numbers against the right edge.
Backend (search_appraisals):
- Add include_nearby + nearby_block_radius flags. When set, fetches three
buckets in one request:
1. exact match (block + plot, original behavior)
2. same block, other plots
3. nearby blocks (block ± radius, default ±2)
- Composite-key dedupe across buckets (gov.il has no stable decision id).
Frontend:
- AppraisalsTable now renders the three buckets as stacked sections with
Hebrew headers ("החלטות תואמות" / "באותו גוש" / "בגושים סמוכים").
- Tab counter sums all three buckets.
Auto-enables include_nearby whenever a block is in scope (address mode
via gush_helka lookup, or block_plot mode directly).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
text-end in dir="rtl" aligns content to the LEFT (end of the line in RTL),
which made numeric columns visually detach from their headers and feel
mis-aligned. Switch all numeric columns to text-start so numbers sit on
the right edge — matching the RTL reading direction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Govmap autocomplete echoes back "גוש N חלקה M" without resolving to a street,
and there's no public reverse-geocoding endpoint. But cadastral parcel
numbers are assigned along street frontage, so the deal with parcelNum
closest to the searched plot (within the same gush) is almost always on
the same street. Use that as the primary signal; fall back to autocomplete
text parsing for address-mode searches.
Example: search gush 6212 plot 894. Plot 896 (הילדסהיימר 12) is the closest
match → "הילדסהיימר" is the searched street → labels work correctly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previous approach: pick most common streetNameHeb among street-source deals.
Problem: a busier nearby street wins even when the searched parcel is on a
quieter street (e.g. רב צעיר beats הילדסהיימר despite being a different street).
Fix: backend now returns resolved_address from Govmap autocomplete (e.g.
"הילדסהיימר 14, ירושלים"). Frontend parses the leading non-numeric tokens
as the street name and uses that as ground truth for proximity labeling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Instead of showing "סביבה קרובה" for all street-polygon deals, we now
infer the searched street from the most common streetNameHeb among
deal_source="street" deals. Deals on that street get "אותו רחוב" (blue);
deals on other streets get "סביבה קרובה" (lighter blue). Works correctly
for both address searches and gush+plot searches.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Add גוש, חלקה, תת-חלקה columns (gushNum/parcelNum/subParcelNum from Govmap)
- Fix house number access: was reading `house_num` (non-existent field),
now reads `houseNum` (actual extra field from Govmap API)
- Rename proximity labels: "באותו רחוב" → "סביבה קרובה", "בשכונה" → "שכונה"
because "street" polygon searches return deals from multiple streets,
making "same street" label factually incorrect
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The deals tab was previously disabled when searching by block+plot
because find_recent_deals_for_address needs an address string. Govmap's
autocomplete in fact accepts the literal "גוש N חלקה M" and returns
the parcel coordinates, so we just compose that string client-side
and let the rest of the pipeline (coords → polygon → deals) run as-is.
Verified live: search "Block=6212 Plot=894" now returns the same two
deals on יהודה המכבי that an address search would.
Appraiser-name search still disables the deals tab — there's no
location to anchor a deals query on.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five-layer defense so that no single library/browser quirk can flip the
layout back to LTR:
1. <body dir="rtl"> in addition to <html dir="rtl"> — some libs read
document.body.dir, not document.documentElement.dir.
2. Explicit `direction: rtl` on html, body in the @layer base CSS.
3. <DirectionProvider dir="rtl"> from @radix-ui/react-direction wrapping
the React tree, so every Radix primitive (Tabs, Dialog, Select,
Popover, Toast) gets RTL context — Radix's useDirection defaults to
"ltr" without it.
4. shadcn Tabs wrapper now defaults its dir prop to "rtl" — redundant
with the provider above, but defends against a future provider
removal silently breaking Tabs.
5. dir="rtl" set explicitly on each <table> element.
Fixes the symptom seen on the live site: deals/appraisals tables had
תאריך (the first <th> in the JSX) rendering on the LEFT instead of the
RIGHT, and Tabs flowed עסקאות→שמאי left-to-right instead of the
expected right-to-left.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The container ran fine (uvicorn started on :8000) but Coolify's healthcheck
runs `curl` / `wget` inside the container, neither of which ships in
python:3.13-slim. Container marked unhealthy → rolled back.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The catch-all `lib/` pattern in both .gitignore and .dockerignore (intended
to exclude Python venv lib dirs) silently swallowed `web/src/lib/utils.ts`.
That's why every Docker build got "ENOENT /web/src/lib/utils" — the file
was never committed and never copied into the container, but local builds
worked because the file existed on disk.
Anchor `lib/` and `lib64/` to the repo root with `/lib/` and `/lib64/`,
then add the previously-ignored utils.ts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Vite's path alias resolution worked locally but consistently failed in
the Docker stage with "Could not load /web/src/lib/utils ... ENOENT"
even after fixing __dirname for ESM. Some interaction between act/buildx
and Vite's resolver around tsconfig paths. Relative imports sidestep the
issue entirely and are explicit about the dependency direction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`__dirname` is undefined in ESM modules, so `path.resolve(__dirname, "./src")`
silently produced a bad path in Docker. Locally it happened to work (some
Node interop), but the Docker stage hit a clear error: "Could not load
/web/src/lib/utils ... ENOENT". Use fileURLToPath(import.meta.url).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Docker build was repeatedly failing on `@/*` path-alias resolution in
tsc, even though local builds (and the same script in the same Node 22
container) succeeded. Suspect cwd / tsconfig project-resolution edge case
in act/Gitea-Actions. Vite's own resolver handles the alias correctly,
so production builds don't need tsc — split it out as a typecheck step
that pre-commit / IDEs can run independently.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The composite-mode + project-references setup broke `@/*` path alias
resolution inside the Docker stage 1 build (worked locally because tsc's
incremental cache was warm). Switch to a single tsconfig with
`tsc --noEmit && vite build` — Vite handles the alias via its own
resolver, so we only use tsc as a type-checker.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pushes to main build a multi-stage Docker image (Node 22 → Python 3.13),
push it to gitea.dev.marcus-law.co.il/mcp-servers/nadlan-mcp:latest
(plus a build-N tag, plus a vN.N.N tag on git tag pushes), then trigger
Coolify to redeploy app b8uf2yysbgeo1un4r9fbwt35.
Mirrors the pattern used by Chaim/decisions-court. Repo secrets set:
REGISTRY_USER, REGISTRY_PASSWORD, COOLIFY_TOKEN.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a self-contained web interface that wraps the MCP tools so end users
don't need to issue MCP commands. Single Docker container serves:
/ → React SPA (RTL Hebrew, Heebo font, shadcn/ui components)
/api/* → FastAPI REST endpoints used by the SPA
/api/docs → OpenAPI / Swagger
/mcp → existing FastMCP streamable-http endpoint
/health → liveness probe
REST endpoints (in nadlan_mcp/web_app.py):
GET /api/search/address?q=... — autocomplete + best-effort gush/חלקה
via Govmap PARCEL_ALL layer
POST /api/search/deals — Govmap deals around an address
POST /api/search/appraisals — Decisive-appraiser search (gov.il)
GET /api/appraisals/pdf?url=.. — Stream PDFs through this server,
carrying the upstream x-client-id
header (a normal browser fetch fails)
UI flow:
- Address search → resolves gush/חלקה → shows BOTH deals (Govmap) AND
appraiser decisions (gov.il) for the same parcel side-by-side in tabs.
- Block+plot search → goes straight to appraisals.
- Appraiser-name search → all decisions by that appraiser.
Stack chosen for RTL-first usability and longevity: React 19, Vite 6,
TypeScript, Tailwind, shadcn/ui (Radix primitives), TanStack Query,
Lucide icons. ~295KB gzipped JS, no SSR overhead.
Multi-stage Dockerfile: Node 22 builds web/ in stage 1, Python 3.13
runtime serves the dist + the FastAPI app in stage 2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The PDF host (free-justice.openapi.gov.il) requires the same x-client-id
header as the search API, so a normal browser click on the URL fails.
This tool carries the auth header automatically and saves the PDF to a
configurable local directory.
Safety properties:
- SSRF guard: only download from free-justice.openapi.gov.il and
pub-justice.openapi.gov.il.
- Path-traversal guard: filename is reduced to its basename; arbitrary
paths are stripped.
- Content-type guard: rejects 200-OK responses whose body is not a real
PDF (gateway sometimes returns JSON-error 200s).
- Atomic write via .tmp + rename so partial downloads never replace the
cached copy.
- Caches by destination path; re-downloads are no-ops unless overwrite=True.
DECISIVE_APPRAISER_DOWNLOAD_DIR env var configures the output dir
(default: ./downloads/decisive_appraisals). 6 new unit tests cover the
happy path, caching, and all four guards. End-to-end live test confirmed
a 1.4MB real PDF lands on disk with valid `%PDF-1.7` header.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a new MCP tool `search_decisive_appraisals` that queries the
Ministry of Justice public registry (~30K published decisions) by
block (גוש), plot (חלקה), appraiser name, committee, decision/publicity
date ranges, or free text — and returns metadata + direct PDF URLs.
Implementation notes:
- New `nadlan_mcp/govil/` package, parallel to `nadlan_mcp/govmap/`,
for gov.il APIs that are not Govmap. Pydantic v2 models match the
upstream PascalCase response via aliases.
- Upstream sits behind an F5 WAF that rejects standard `requests`;
uses `curl_cffi` with Chrome 120 impersonation to traverse it.
- Static `x-client-id` header (issued to the gov.il SPA, public, visible
in any DevTools session) is required by the gateway — without it
every call returns a generic 500.
- 14 unit tests cover model parsing, body shape, pagination, and the
500-is-fatal contract (configuration error, not retryable).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Feature:
- New parameter `include_outlier_deals` (default=True) in: - filter_deals_for_analysis()
- calculate_deal_statistics()
- get_valuation_comparables() MCP tool
- get_deal_statistics() MCP tool
Behavior:
- When ON: response includes `outlier_deals` field with removed deals
- LLM can see what was filtered out, answer questions about it
- Maintains full transparency on outlier removal
Terminology fixed:
- "outlier_deals" = deals removed as outliers (clearer than "filtered_deals")
- "filtered deals" = deals that PASSED filtering
- Added to OutlierReport model
All 311 tests pass
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Replaced runtime checks with decorator-based registration:
- Created conditional_tool() decorator
- Disabled tools not registered with FastMCP at all
- Client queries won't see disabled tools
- Removed "This tool is currently disabled" messages
Disabled tools (invisible to clients):
- get_deals_by_radius
- get_street_deals
- get_market_activity_metrics
All 311 tests pass, 17 skipped
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Bug #1: Fixed IndexError in outlier_detection.py:256
- Missing enumerate() caused stale loop var to access beyond bounds
- Occurred when hard bounds filtered deals before IQR processing
- Added test reproducing exact scenario (21 deals → 8 filtered)
Bug #2: Added stack traces to all MCP tool error logs
- Added exc_info=True to 10 MCP tools' error handlers
- Improves debugging by logging full stack traces
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Rationale: 50% threshold too permissive for Israeli real estate market.
Within same room count/area, legitimate price variance typically ±30-35%, not 50%.
Example: 13,402 NIS/sqm deal (47% below 25,120 median) was passing with 50% threshold.
This is almost certainly data error or partial deal, not legitimate market variation.
Changes:
- ANALYSIS_PERCENTAGE_THRESHOLD default: 0.5 → 0.4 (40%)
- Update CLAUDE.md docs to reflect 40% threshold
Impact:
- New lower bound: median × 0.6 (was median × 0.5)
- New upper bound: median × 1.4 (was median × 1.5)
- Tighter filtering while preserving legitimate high/low-end deals
- Better aligned with actual Israeli real estate market variance
Tests: All 326 tests pass
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Change ANALYSIS_IQR_MULTIPLIER default from 1.5 to 1.0 in config.py
- Add iqr_multiplier parameter to all filtering & statistics functions
- Allow runtime override via MCP tools (get_valuation_comparables, get_deal_statistics)
- Update CLAUDE.md docs with new default & override examples
Rationale: k=1.0 catches more suspicious deals (e.g. 43% below median) while
still preserving legitimate edge cases via hard bounds. Users can override
per-call for more conservative filtering (k=1.5) if needed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Adds visibility into the deal filtering pipeline to help diagnose
why certain queries return fewer deals than expected.
New logs show:
1. Number of deals before criteria filtering
2. Number of deals after criteria filtering (with count removed)
3. Whether outlier filtering was applied
4. Number of deals after outlier filtering (with count removed)
5. Outlier filtering method and parameters used
Example output:
INFO: Applying criteria filters to 41 deals
INFO: After criteria filtering: 12 deals (removed 29 deals)
INFO: After outlier filtering (iqr, k=1.5): 4 deals (removed 8 outliers)
This helps users and developers understand:
- If criteria filters are too restrictive
- If outlier filtering is too aggressive
- Where deals are being filtered out in the pipeline
Addresses user question: "Why do I get just 4 deals for a central address
like סירקין 16 תל אביב?" - Now they can see exactly where deals were
filtered out.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Bug: When the Govmap API returns the same polygon_id multiple times
(which can happen for addresses with multiple associated polygons),
the code was processing each duplicate separately, leading to:
- Duplicate API calls for street/neighborhood deals
- Wasted API quota and slower performance
- Confusing logs showing same polygon queried multiple times
Example from logs:
INFO: Querying polygon 53283601 (distance: 0m from search point)
INFO: Getting street deals for polygon: 53283601
INFO: Querying polygon 53283601 (distance: 0m from search point)
INFO: Getting street deals for polygon: 53283601
Root cause:
- nearby_polygons API response can contain duplicate polygon_ids
- Code was appending all polygons without deduplication
- Log message claimed "unique polygon IDs" but didn't enforce it
Solution:
- Use dict to deduplicate polygons by polygon_id (lines 612-643)
- When duplicates exist, keep the one with shortest distance
- Convert dict to list after deduplication
- Now truly have unique polygon IDs as the log claims
Impact:
- Reduces API calls (fewer duplicates = less load on Govmap API)
- Faster queries (less redundant processing)
- Clearer logs (no confusing duplicate entries)
- Better performance for addresses with many associated polygons
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The test_market_activity_trend_stable test was failing intermittently
because it used get_recent_date() which approximates months as 30 days,
causing inconsistent month boundary calculations depending on when the
test runs.
Root cause:
- Using months_ago * 30 creates shifting month boundaries
- Calendar dates don't align consistently with 30-day periods
- Deals could fall into unexpected months based on test execution date
- Trend calculation compares first half vs second half of months
- Inconsistent month grouping led to "decreasing" instead of "stable"
Solution:
- Use explicit, deterministic dates (15th of each month)
- Calculate year-month combinations going back from current month
- Ensures exactly 1 deal per month for 12 consecutive months
- All deals guaranteed to be within time_period_months=12 window
- Month boundaries are now consistent regardless of test execution date
Before: Test result varied based on current date (flaky)
After: Test result is always "stable" (deterministic)
This ensures the test validates the trend detection logic without
temporal flakiness.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The MCP now includes detailed outlier filtering information in the response,
allowing LLMs and bots to inform users about data quality improvements.
Changes to nadlan_mcp/fastmcp_server.py:
- Updated get_valuation_comparables to include outlier filtering metadata
- deal_breakdown now includes when filtering is applied:
- total_deals: Count after filtering (final result)
- total_deals_before_filtering: Original count before filtering
- outliers_removed: Number of deals filtered out
- filtering_method: Method used ("iqr", "percent", or "none")
- iqr_multiplier: IQR multiplier when using IQR method (e.g., 1.5)
- Updated docstring to clarify that outlier filtering is automatic
and metadata is included in response
Changes to tests/e2e/test_mcp_tools_comprehensive.py:
- Added assertions to verify outlier filtering metadata fields
- Test now validates the structure and types of filtering information
Example response:
{
"market_statistics": {
"deal_breakdown": {
"total_deals": 15,
"total_deals_before_filtering": 17,
"outliers_removed": 2,
"filtering_method": "iqr",
"iqr_multiplier": 1.5
}
}
}
This allows bots like nadlan-bot to properly inform users:
"Found 17 comparable deals, filtered 2 outliers using IQR method (k=1.5),
showing 15 deals."
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Added log_mcp_call() helper function that logs all MCP tool invocations
with their parameters at INFO level. This provides:
- Better debugging and monitoring capabilities
- Clear visibility into which tools are being called and with what parameters
- Smart parameter formatting (truncates long strings, summarizes long lists)
- Consistent logging format across all 10 MCP tools
Example log output:
INFO - MCP tool called: find_recent_deals_for_address(address=סוקולוב 38 חולון, years_back=2, radius_meters=30, max_deals=100, deal_type=2)
This helps track API usage patterns and debug issues in production.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The same-building detection was failing because the code was accessing
deal.street_name and deal.house_number (Pydantic model field names),
but the Govmap API actually returns streetNameHeb/streetNameEng and
houseNum as extra fields.
Modified address construction to check multiple field names using
getattr() to handle the API's actual field names, falling back to
model field names if needed.
This fix ensures same-building deals are correctly identified and
prioritized with priority=0.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Implements smart deal prioritization based on distance from search point,
ensuring most relevant comparables are selected first.
Problem:
- Queried all polygons equally without distance consideration
- No filtering for deals that are too far away
- Street/neighborhood deals could be from distant locations
- No way to prefer closest polygon when multiple available
Solution:
1. Sort polygons by distance from search point (closest first)
2. Query polygons progressively, stop early if enough deals found
3. Calculate and store distance_meters for each deal
4. Filter deals by configurable distance thresholds:
- Street deals: max 500m (configurable)
- Neighborhood deals: max 1000m (configurable)
5. Sort by priority → distance → date (prefer closer deals)
Changes:
Config (nadlan_mcp/config.py):
- Added max_street_deal_distance_meters (default: 500m)
- Added max_neighborhood_deal_distance_meters (default: 1000m)
Client (nadlan_mcp/govmap/client.py):
- Extract polygon metadata with coordinates
- Sort polygons by distance using utils.calculate_distance()
- Progressive polygon querying (closest first, stop when enough deals)
- Calculate deal distance (use polygon distance as approximation)
- Filter street deals beyond max_street_deal_distance_meters
- Filter neighborhood deals beyond max_neighborhood_deal_distance_meters
- Store distance_meters on each Deal (dynamic attribute)
- Updated sorting: Priority → Distance → Date
Benefits:
- More relevant comparables (from nearby locations)
- Reduced API calls (query fewer polygons, stop early)
- Better performance (fewer deals to process)
- Configurable distance thresholds
- Transparent (distance info available in each deal)
Example Configuration:
export MAX_STREET_DEAL_DISTANCE_METERS=300 # Conservative
export MAX_NEIGHBORHOOD_DEAL_DISTANCE_METERS=500
Testing:
- All 326 tests pass
- Backward compatible (high default limits)
- Works with existing outlier filtering
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Problem:
- test_get_market_activity_metrics was failing with "list index out of range"
- _safe_calculate_metric only caught ValueError, not IndexError or other exceptions
- This caused the entire function to crash when metric calculations failed
Root Cause:
- Market metric functions (calculate_market_activity_score, get_market_liquidity,
analyze_investment_potential) can throw IndexError if insufficient data
- The helper function wasn't catching these exceptions
Solution:
- Changed _safe_calculate_metric to catch all exceptions (Exception instead of ValueError)
- Added logging to track which metric failed
- Now returns error dict gracefully instead of crashing
Impact:
- get_market_activity_metrics is now more robust
- Returns partial results even if some metrics fail
- All 326 tests now pass
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Problem:
- get_valuation_comparables calculated statistics with outlier filtering
- But returned a deals list that still contained outliers (e.g., ₪900K, ₪1.3M deals)
- This caused inconsistency between statistics and the actual deals shown
Root Cause:
- calculate_deal_statistics() filters outliers internally for statistics
- But the deals array returned to user was only filtered by criteria (rooms, price, etc.)
- Outlier filtering was not applied to the returned deals list
Solution:
- Added explicit call to filter_deals_for_analysis() after filter_deals_by_criteria()
- Now both statistics AND deals list use outlier-filtered data
- Moved imports to top of file for better practice
Changes:
- nadlan_mcp/fastmcp_server.py:
- Added imports: get_config, filter_deals_for_analysis
- In get_valuation_comparables(): Apply outlier filtering before calculating stats
- Ensures deals list and statistics are consistent
Testing:
- 325/326 tests pass
- 1 unrelated test failure in get_market_activity_metrics (pre-existing issue)
After deployment, outliers will be automatically filtered from valuation comparables.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The .dockerignore was excluding pyproject.toml, which is needed for
`pip install .` to work after the project was modernized to use
pyproject.toml instead of setup.py.
This was causing Docker builds to fail with:
"failed to calculate checksum: /pyproject.toml: not found"
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Implement configurable outlier detection and robust statistical measures to
improve analysis accuracy for real estate data. Addresses issues with data
entry errors, partial deals, and other anomalies that skew statistics.
Key Features:
- IQR-based outlier detection (moderate filtering by default, k=1.5)
- Hard bounds filtering for obvious errors (price_per_sqm, deal_amount)
- Robust volatility using IQR instead of std_dev for investment analysis
- Transparent reporting with both filtered and unfiltered statistics
Implementation:
- Add outlier_detection.py module with IQR/percent/hard bounds methods
- Add OutlierReport model and enhance DealStatistics with filtered fields
- Update calculate_deal_statistics() to support optional outlier filtering
- Update analyze_investment_potential() to use robust volatility
- Add 9 new configuration parameters for customization
- Add comprehensive test suite (24 tests) for outlier detection
- Update CLAUDE.md with usage documentation
Configuration (all via env vars):
- ANALYSIS_OUTLIER_METHOD=iqr (default, or percent/none)
- ANALYSIS_IQR_MULTIPLIER=1.5 (moderate, 3.0=conservative)
- ANALYSIS_PRICE_PER_SQM_MIN/MAX=1000/100000 (bounds in NIS/sqm)
- ANALYSIS_MIN_DEAL_AMOUNT=100000 (catches partial deals)
- ANALYSIS_USE_ROBUST_VOLATILITY=true (IQR-based CV)
- ANALYSIS_USE_ROBUST_TRENDS=true (filter before regression)
Testing:
- All existing tests pass (326 passed)
- 24 new comprehensive outlier detection tests
- Real-world scenario tests (partial deals, data errors)
Backward Compatible:
- Default behavior improves accuracy without breaking changes
- All new fields in models are optional
- Config parameters have sensible defaults
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Remove requirements.txt and requirements-dev.txt (duplicates pyproject.toml)
- Update Dockerfile to use 'pip install .' instead of requirements.txt
- Update README.md and GitHub workflows to use 'pip install -e .[dev]'
- Fix test_get_valuation_comparables: check for 'deal_date' instead of non-existent 'asset_room_num'
(rooms field is optional and excluded when None due to exclude_none=True)
- Fix fastmcp version constraint in pyproject.toml (>=2.13.0,<3.0.0)
- Add vcrpy to dev dependencies
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The Deal model's 'rooms' field was missing the 'assetRoomNum' alias,
causing room data from the API to not be loaded into the model.
This resulted in all room-based filters returning 0 results.
Bug discovered when querying "חנקין 62 חולון" for 3-room apartments.
Query returned 0 results despite multiple 3-room deals existing in the data.
Root cause: Pydantic requires field aliases to map API field names (camelCase)
to Python attributes (snake_case). The 'rooms' field had no alias.
Fix: Added alias="assetRoomNum" to rooms field in Deal model.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Enables deployment to cloud platforms (Render, Railway, etc.) while maintaining
backward compatibility with existing stdio transport for Claude Desktop.
New features:
- HTTP server entry point (run_http_server.py) using uvicorn
- Docker containerization with Python 3.13
- Health check endpoint at /health
- Comprehensive deployment documentation for Render, Railway, and Docker
Technical changes:
- Added uvicorn dependency for ASGI server
- Created Dockerfile with optimized multi-stage build (343MB)
- Added .dockerignore for efficient Docker builds
- Implemented /health endpoint using Starlette JSONResponse
- Updated README.md and DEPLOYMENT.md with HTTP deployment guides
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
chaim
Nitzan Pomerantz