AskSG

A Retrieval-Augmented Generation (RAG) assistant for Singapore public policy documents.

Ask questions in natural language about Singapore's Budget, housing policies, CPF rules, and monetary policy — and get answers grounded in official government sources.

---

What It Does

AskSG lets you ask questions like:

• "What did Budget 2025 say about support for first-time homebuyers?"
• "Am I eligible to buy an HDB resale flat as a Singapore PR?"
• "How has Singapore's fiscal policy stance changed from 2023 to 2025?"
• "What are the CPF contribution rates for employees above 55?"
• "What is MAS's current assessment of Singapore's inflation outlook?"

Instead of reading through long government PDFs, you get a cited, grounded answer in seconds.

---

Architecture

Singapore government documents
   (Budget speeches, MAS reviews, HDB guides, CPF guides)
            │
            ▼
    [ Ingestion ]                pipelines/ingest_documents.py
    Download PDFs + HTML
            │
            ▼
    [ Preprocessing + Indexing ] pipelines/build_indexes.py
    Clean + split into
    overlapping text chunks
    (500 chars, 50 overlap)
    → ChromaDB vector store
    → BM25 keyword index
            │
            ▼
    [ Retrieval ]                src/retrieval/
    User question
    → embed query (MiniLM)
    → BM25 keyword retrieval  ─┐
    → dense vector retrieval  ─┴─ RRF fusion (top-25)
    → cross-encoder rerank (top-9)
            │
            ▼
    [ Generation ]               src/generation/answer.py
    → Groq LLM (Llama 3.3 70B)
    → grounded answer
            │
            ▼
    [ Streamlit Interface ]      app/main.py
    Chat interface

---

Data Sources

AskSG draws from 25 official Singapore government documents across 6 source types, totalling ~762K characters of extracted policy text.

Source	Documents	Format	Topics covered
Singapore Budget	Budget speeches 2023, 2024, 2025, 2026	PDF	Fiscal policy, grants, cost-of-living transfers, CPF and HDB enhancements
MAS	Macroeconomic Reviews Oct 2024, Apr 2025 (each including the Monetary Policy Statement)	PDF	Monetary policy statements, GDP and inflation forecasts, labour market, exchange rate policy
HDB	Eligibility guides — couples & families; singles	JS-rendered HTML	BTO/resale eligibility rules, CPF housing grants, income ceilings, SPR household rules
CPF Board	Contribution rates, housing usage, OW ceiling, Budget 2023 highlights, retirement sums (BRS/FRS/ERS), milestones at 55 and 65, CPF LIFE, retirement withdrawals, CPFIS options, Medisave usage, Medisave BHS, interest rates, retirement top-up scheme, top-up tax relief	HTML	Full CPF lifecycle — contributions, housing, retirement planning, Medisave, interest rates, investment scheme
IRAS	SRS tax relief overview	HTML	Supplementary Retirement Scheme — contribution limits, tax relief, withdrawal rules
MAS	Singapore Savings Bonds overview	HTML	SSB features, step-up interest structure, application and redemption rules

All documents are official Singapore government publications, freely accessible to the public.

Why these sources?

Budget speeches (2023–2026): The annual Budget is the primary vehicle for fiscal policy announcements — grants, CPF rule changes, HDB grant amounts, cost-of-living transfers, and tax adjustments. Four consecutive years gives the system temporal context so it can distinguish what was announced in which year, and answer comparison questions across budgets.

MAS Macroeconomic Reviews: Published twice yearly, each review bundles the Monetary Policy Statement (MPS) with MAS's full assessment of Singapore's growth, inflation, labour market, and external outlook. Apr 2025 is the most recent and contains 2025 growth and inflation forecasts. Oct 2024 provides the prior half-year policy context and baseline.

Note

MAS special features trimming — This is the only document in the corpus that was partially extracted rather than used in full. The Oct 2024 review is a 100+ page academic publication. Only the main policy chapters (MPS, economic outlook, inflation, monetary and fiscal policy, data boxes) were kept. The three academic Special Feature articles — Proceedings of the 2024 Asian Monetary Policy Forum, A Perspective on Inflation Targeting, and Globalisation is Not Dying, It's Just Changing — were removed. These articles add ~69K chars of off-topic research content that would compete with policy chunks in retrieval without answering any realistic user query.

HDB eligibility guides: HDB's website is JavaScript-rendered and required Playwright for content extraction. The couples & families page (59K chars) covers eligibility by household type, CPF housing grants (Enhanced, Proximity, Step-Up), income ceilings, and the SPR 3-year rule for SPR-only households buying resale flats. The singles page (21K chars) covers the Single Singapore Citizen scheme. HDB redesigned their site in 2025 — old URL paths now return a "We have moved" page; ETL uses the current /flat-grant-and-loan-eligibility/ paths.

CPF guides (15 pages): CPF is the most common topic for Singapore personal finance questions. Coverage spans the full CPF lifecycle: contribution rates by age band, OA savings for housing purchases, retirement sum milestones (BRS $106,500 / FRS $213,000 / ERS $426,000 for 2025), CPF LIFE monthly payouts, what happens at 55 and 65, retirement withdrawal rules, CPFIS for investing OA/SA savings, Medisave usage and the Basic Healthcare Sum ($79,000 for 2026), current interest rates (OA 2.5% / SA-MA-RA 4%), and the Retirement Top-Up Scheme (up to $16,000 tax relief per year).

SRS (IRAS): The Supplementary Retirement Scheme lets members voluntarily top up savings for additional tax relief — up to $15,300/year for Singapore Citizens and PRs, $35,700 for foreigners. Covers contribution caps, eligible SRS investments, and the withdrawal rules at retirement age (63).

Singapore Savings Bonds (MAS): Risk-free, flexible government bonds with step-up interest rates backed by the Singapore Government. The MAS page explains the application process, the $200,000 individual limit, and no-penalty early redemption — a common question from risk-averse savers.

Corpus composition

Source	Files	Chars	Share	Period covered
Budget	4	336,610	44.1%	FY2023 – FY2026 (4 annual speeches)
MAS	2	240,048	31.5%	Oct 2024 – Apr 2025 (2 half-yearly reviews, each including the MPS)
CPF	15	89,502	11.7%	2023 – 2026 (salary ceiling from Budget 2023; interest rates, BHS current as of 2026)
HDB	2	79,836	10.5%	Current as of May 2026
SRS	1	9,185	1.2%	Current as of May 2026
SSB	1	6,976	0.9%	Current as of May 2026
Total	25	762,118

Chunks are 500-character overlapping windows with 50-char overlap, generated by pipelines/build_indexes.py.

---

Tech Stack

Layer	Tool
Language	Python 3.9+
Embeddings	sentence-transformers — all-MiniLM-L6-v2
Keyword index	rank-bm25 — BM25Okapi over all corpus chunks
Vector store	ChromaDB (local, persistent)
Reranker	cross-encoder/ms-marco-MiniLM-L-6-v2
LLM	Groq API — llama-3.3-70b-versatile (free tier)
PDF parsing	pdfplumber
Frontend	Streamlit
Deployment (planned)	FastAPI + Docker + AWS EC2

---

Project Structure

asksg/
│
├── configs/                             # What to fetch and how to configure retrieval
│   ├── sources/
│   │   ├── document_sources.py          # Policy documents registry (Budget, CPF, HDB, MAS, SRS, SSB)
│   │   └── structured_sources.py        # HDB resale transaction dataset (data.gov.sg API)
│   └── retrieval/
│       ├── baseline.yaml                # Dense-only: embed_model, top_k
│       ├── hybrid.yaml                  # BM25 + dense + RRF: fetch, rrf_k, top_k
│       └── reranker.yaml                # Full pipeline: adds cross-encoder model + top_n
│
├── data/                                # All data artefacts (partially git-tracked)
│   ├── raw/                             # Downloaded files — not tracked
│   │   ├── documents/                   # PDF / HTML snapshots
│   │   └── structured/                  # Raw HDB resale API exports
│   ├── interim/
│   │   ├── extracted_text/              # Parsed .txt files — tracked in git
│   │   │   ├── budget/                  # Budget speeches 2023–2026 (4 PDFs)
│   │   │   ├── mas/                     # MAS Macroeconomic Reviews (2 PDFs, each includes MPS)
│   │   │   ├── hdb/                     # HDB eligibility guides (2 JS-rendered HTML)
│   │   │   ├── cpf/                     # CPF guides — full lifecycle (15 HTML)
│   │   │   ├── srs/                     # IRAS SRS overview (1 HTML)
│   │   │   └── ssb/                     # MAS Savings Bonds overview (1 HTML)
│   │   └── cleaned_tables/              # Typed HDB resale CSV — not tracked
│   ├── processed/
│   │   └── chunks.jsonl                 # 500-char overlapping chunks — tracked in git
│   └── indexes/
│       └── chroma/                      # ChromaDB vector index — not tracked (rebuilt via make index)
│
├── src/                                 # All library code — import as src.*
│   ├── ingestion/
│   │   ├── documents/                   # Unstructured document pipeline
│   │   │   ├── fetchers/
│   │   │   │   ├── http_fetcher.py      # HTTP transport: returns raw bytes (PDF + HTML)
│   │   │   │   └── playwright_fetcher.py# JS-rendered pages: headless Chromium → raw HTML bytes
│   │   │   ├── extractors/
│   │   │   │   ├── pdf_extractor.py     # pdfplumber: bytes → plain text
│   │   │   │   └── html_extractor.py    # BeautifulSoup: bytes → plain text
│   │   │   └── pipeline.py             # Orchestrates: source registry → fetch → extract → save
│   │   └── structured_data/             # Structured data pipeline (Agentic RAG input)
│   │       ├── hdb_api_fetcher.py       # Paginated data.gov.sg CKAN API client
│   │       └── schemas.py               # Column definitions and expected types
│   ├── processing/
│   │   ├── clean_text.py                # Normalise whitespace, strip PDF artefacts
│   │   ├── chunker.py                   # Recursive character splitter with overlap
│   │   └── clean_hdb_resale.py          # Type coercion + derive year/month columns
│   ├── indexing/
│   │   ├── build_vector_index.py        # Embed chunks → ChromaDB (all-MiniLM-L6-v2)
│   │   └── build_keyword_index.py       # Build BM25Okapi from chunks.jsonl
│   ├── retrieval/
│   │   ├── loader.py                    # Load all components once per session (NamedTuple)
│   │   ├── dense.py                     # Baseline: vector search only
│   │   ├── bm25_retriever.py            # Keyword search via BM25Okapi
│   │   ├── hybrid.py                    # BM25 + dense fused with Reciprocal Rank Fusion
│   │   ├── reranker.py                  # Cross-encoder reranking (ms-marco-MiniLM-L-6-v2)
│   │   └── metadata_filter.py          # Dense-side source/year filter for targeted queries
│   ├── generation/
│   │   ├── prompts.py                   # SYSTEM_PROMPT + context block formatter
│   │   └── answer.py                    # answer() blocking + stream_answer() streaming
│   └── tools/                           # Agentic RAG tools (stubs — Phase 2)
│       ├── policy_search.py             # Wraps hybrid+rerank retrieval as agent-callable tool
│       └── hdb_resale_query.py          # Structured query over HDB resale parquet
│
├── pipelines/                           # CLI entry points — orchestrate src/* modules
│   ├── ingest_documents.py              # Fetch + extract all policy documents
│   ├── ingest_hdb_data.py               # Fetch HDB resale CSV + clean to parquet
│   ├── build_indexes.py                 # Chunk documents + build vector and keyword indexes
│   ├── run_eval.py                      # Run RAG evaluation (hit rate, MRR, evidence recall, answer metrics)
│   └── extract_local_pdf.py             # Fallback: extract text from a manually downloaded PDF
│
├── experiments/                         # Ablation study runs — one file per pipeline variant
│   ├── baseline.py                      # Dense-only retrieval (Stage 1 baseline)
│   ├── hybrid_retrieval.py              # BM25 + dense + RRF, no reranker
│   ├── with_reranker.py                 # Full pipeline: hybrid + cross-encoder rerank
│   └── compare_results.py              # Print comparison table across eval/results/
│
├── app/
│   └── main.py                          # Streamlit chat interface
│
├── eval/
│   ├── ragas_eval.py                    # Two-tier eval: local NLI + LLM-as-judge
│   ├── test_set.json                    # Curated Q&A pairs with ground truth
│   └── results/                         # Per-run eval output JSON files
│
├── Makefile                             # make fetch / index / app / baseline / hybrid / reranker
├── requirements.txt
└── README.md

---

Getting Started

# 1. Clone and set up environment
git clone https://github.com/isafzh/asksg.git
cd asksg
python -m venv .venv
.venv\Scripts\activate       # Windows
# source .venv/bin/activate  # Mac/Linux
pip install -r requirements.txt

# 2. Add your Groq API key
echo GROQ_API_KEY=your_key_here > .env
# Get a free key at https://console.groq.com

# 3. Build the vector index (downloads ~90MB model on first run)
# Windows/OneDrive: set this BEFORE building — Chroma's SQLite fails under OneDrive sync
$env:ASKSG_CHROMA_DIR = "$env:LOCALAPPDATA\AskSG\chroma"   # PowerShell (run once, or: setx ASKSG_CHROMA_DIR ...)
make index
# Windows without make: py pipelines\build_indexes.py

# 4. Run the app
make app
# → opens at http://localhost:8501

To refresh the document corpus:

make fetch      # re-download source documents
make clean      # delete old index
make index      # re-chunk and re-embed

---

Evaluation

RAG quality is measured on a hand-curated test set of 30 Q&A pairs drawn from the source documents (eval/test_set.json).

Test set composition

Domain	How many	Which questions	Rationale
CPF	6	Q1, Q2, Q3, Q19, Q20, Q21	Most rule-heavy domain; tests contribution, housing, retirement, and LIFE rules
Budget	5	Q4, Q5, Q6, Q22, Q23	Temporal disambiguation and year-specific policy announcements
HDB	4	Q7, Q8, Q17, Q18	Core housing eligibility and grant questions; HDB also has a planned structured transaction-data modality for the agentic tool
MAS	3	Q9, Q10, Q24	Macro forecasts and monetary policy framework; tests numeric projection retrieval and policy-mechanism explanation
SRS	3	Q11, Q12, Q13	Smaller but distinct retirement/tax-relief source
SSB	3	Q14, Q15, Q16	Smaller but distinct investment product source
Cross-domain	6	Q25–Q30	Cross-domain topics spanning two policy areas; four require evidence from two domains, while two are domain-overlap questions answerable from one expected document
Total	30

Cross-domain pairs:

Q25–Q27, Q29 require evidence from two corpus buckets:

• Q25 srs+cpf — comparing SRS contributions vs CPF RA top-ups for tax relief
• Q26 ssb+srs — using SRS funds to invest in SSBs, and how returns are taxed
• Q27 budget+mas — MAS's April 2025 GDP projection vs Budget 2026's reported actual 5% outcome
• Q29 cpf+budget — CPF and Budget measures for lower-income retirement adequacy

Q28, Q30 are domain-overlap questions answerable from one expected document:

• Q28 hdb+budget — housing options and grants for singles buying near parents (evidence: HDB eligibility doc)
• Q30 cpf+hdb — CPF OA usage when buying an HDB flat, plus refund obligation when sold (evidence: CPF housing usage doc)

Test question schema

Each question carries structured metadata supporting three levels of retrieval evaluation:

Level	What it checks	Schema field
L1 — Source	Was the right document in the top-k results?	expected_sources[].document
L2 — Evidence	Do the retrieved chunks contain the key facts verbatim?	must_contain[]
L3 — Chunk	Was the exact chunk retrieved?	deferred — expected_chunks added after chunking strategy is locked

must_contain strings are grepped from the actual .txt corpus files so they match the exact character sequences the chunker indexes (e.g., tables may store 37 not 37%; strings are stored as ASCII so range notation uses a hyphen - not an en-dash).

Questions are additionally tagged:

Field	Values
difficulty	standard · temporal · multi_chunk · cross_domain
answer_type	numeric · policy_fact · eligibility · comparison
retrieval_mode	unstructured (all 30 current questions) · structured (reserved for future HDB resale data queries)

domain describes the policy area(s) a question touches. difficulty describes what makes retrieval hard. Therefore, a question can be cross-domain by topic but not have difficulty: cross_domain if the required evidence is contained in one expected document.

I initially evaluated the Ragas library, but it required 1,300+ LLM calls per run and exhausted the Groq free-tier quota (100k tokens/day) in a single run. This is a cost-aware RAG evaluation suite inspired by the RAG triad: retrieval quality, groundedness, and answer relevance.

Layer	Metric	What it measures	Method	API cost
Retrieval	Hit Rate@K	Was the expected document in the top-K results?	expected_sources match vs chunk metadata	0 extra calls
Retrieval	MRR@K	How high was the first correct source ranked?	Reciprocal rank of first hit	0 extra calls
Retrieval	Evidence Recall	Do retrieved chunks contain the required facts?	must_contain strings present in context	0 extra calls
Retrieval	Context Relevance	Are retrieved chunks on-topic?	Cosine similarity: query vs chunks (MiniLM)	0 extra calls
Answer	Answer Fact Recall	Does the answer state the required facts?	must_contain strings present in answer	0 extra calls
Answer	Answer Similarity	Does the answer match ground truth semantically?	Cosine similarity: answer vs ground truth (MiniLM)	0 extra calls
Answer	Faithfulness (NLI)	Is every answer claim entailed by the context?	NLI cross-encoder (nli-deberta-v3-base, local)	0 extra calls
Answer	Faithfulness (LLM)	Is every claim grounded? (multi-chunk synthesis)	LLM-as-judge structured prompt	1 call/question
Answer	Answer Relevance (LLM)	Does the answer address the question?	LLM-as-judge structured prompt	1 call/question

Total per full run: 40 Groq API calls (30 generation + 10 LLM judge on a fixed balanced sample covering all 6 source domains, all 4 difficulty levels, and all 4 answer types). The original Ragas library required 1,300+ calls per run — this implementation achieves the same conceptual coverage at 3% of the cost. Use --judge-sample 0 to skip the LLM judge entirely for free iteration.

Baseline Benchmark — Dense-only retrieval (30 questions)

Dense-only retrieval establishes a reference point and reveals failure modes. It is not a construction stage of the final system — it exists to prove the final pipeline improves something measurable.

Metric	k=5	k=7
Hit Rate@K	0.8333	0.9000
MRR@K	0.6917	0.7020
Evidence Recall	0.5278	0.6111
Answer Similarity	0.8082	0.8284
Faithfulness (NLI)	0.4015	0.4403

Hit rate and evidence recall improve with k, but MRR grows only slightly — the right document enters the retrieved set but is not consistently ranked first. Slow MRR growth signals a ranking weakness that expanding k alone cannot fix: dense embeddings treat year-agnostic same-topic chunks as equally relevant regardless of the queried year.

Best completed baseline setting: k=7 — used as the comparison point for the retrieval upgrade below.

Retrieval Upgrade — Hybrid search and reranking (30 questions)

Motivation: baseline MRR is 0.70 at k=7, and temporal disambiguation fails — dense retrieval ranks Budget 2023/2024/2025 chunks as equally relevant for year-specific queries. Fix: BM25 keyword scoring pulls year-specific documents into the candidate pool via RRF fusion; the cross-encoder reranker optionally re-scores the fused candidates before generation.

Why k-sweep here? In the hybrid+rerank pipeline, the initial candidate pool is fixed (top-25 from RRF fusion). top_k controls how many post-rerank chunks are passed to the LLM — a context-size decision, not a retrieval-depth decision. A retrieval-only k-sweep selects the optimal context window without burning Groq quota.

Retrieval-only k-sweep (0 Groq calls):

k	Hit Rate	MRR	Evidence Recall	Context Relevance
5	0.9667	0.8033	0.6389	0.6392
7	0.9667	0.8033	0.6889	0.6246
9	0.9667	0.8033	0.7389	0.6132

Hit Rate and MRR are flat across all k — the reranker consistently surfaces the right document regardless of pool size. Evidence recall improves +5pp per step; context relevance falls slightly as less-semantically-tight supporting chunks are added. k=9 chosen: maximises evidence recall (73.9%) with an acceptable context relevance trade-off.

With dense-side metadata filtering subsequently added (constraining Chroma dense retrieval by source/year for targeted queries), k=9 further improves: Hit Rate 1.0000 / MRR 0.8283 / Evidence Recall 0.7722 / Context Relevance 0.6075.

Reranker ablation — hybrid-only vs hybrid+reranker at k=9 (retrieval metrics):

Metric	Hybrid (no reranker)	Hybrid + Reranker	Δ
Hit Rate@9	0.9667	0.9667	0.0
MRR@9	0.8025	0.8033	+0.0008
Evidence Recall	0.7389	0.7389	0.0
Context Relevance	0.6154	0.6134	–0.002

The reranker adds no measurable retrieval benefit on this corpus at k=9. Hit Rate and Evidence Recall are identical; MRR difference (0.0008) is within noise. The reranker is an optional precision layer — it is included in the final pipeline as a robustness measure but its contribution on these 30 questions is negligible. On larger or more ambiguous corpora the benefit would likely be more pronounced.

Final result — hybrid_rerank+filter k=9 (30 questions):

Layer	Metric	Best dense baseline (k=7)	Hybrid+Rerank+Filter k=9	Δ
Retrieval	Hit Rate@K	0.9000	1.0000	+10.0pp
Retrieval	MRR@K	0.7020	0.8283	+12.6pp
Retrieval	Evidence Recall	0.6111	0.7722	+16.1pp
Retrieval	Context Relevance	—	0.6075	—
Answer	Answer Fact Recall	—	0.5944	—
Answer	Answer Similarity	0.8284	0.8471	+1.9pp
Answer	Faithfulness (NLI)	0.4403	0.3671	–7.3pp
Judge (n=10)	Faithfulness (LLM)	—	0.9600	—
Judge (n=10)	Answer Relevance (LLM)	—	0.9200	—

Metrics from hybrid_rerank_k9.json, the final post-filter judged run over all 30 questions.

NLI faithfulness (0.37) lags the LLM judge (0.96) — the NLI cross-encoder penalises paraphrase even when the claim is factually supported. The LLM judge is the more reliable faithfulness signal.

The sampled LLM judge is used as a grounding/relevance diagnostic, not as the sole correctness metric. Q27 shows why: the answer can be faithful to retrieved context while still missing required benchmark facts — which is captured by evidence_recall and answer_fact_recall, not by the judge. This is intentional: the metrics cover different failure modes.

Q22 partial improvement: Metadata filtering retrieves the Budget 2026 source (hit=1.0, mrr=0.200), but only one of the two required facts appears in the top-9 context (evidence_recall=0.5). The second Budget 2026 chunk ranks below position 9 after reranking.

Q27 chunk-level evidence failure: The expected document is retrieved (hit=1.0, mrr=1.0), but the required evidence string is absent from the top-9 chunks (evidence_recall=0.0, answer_fact_recall=0.0). The LLM judge still scores faithfulness=1.0 and answer_relevance=1.0 because the generated answer is grounded in the available context and sounds relevant — it just cannot state the specific fact that was not retrieved. This is a chunk-level retrieval gap, not a corpus gap (the full document is in the index; the required sentence did not surface in the top-9).

To run:

make baseline   # dense-only benchmark              → eval/results/baseline_k7.json
make hybrid     # reranker ablation, retrieval-only  → eval/results/hybrid_k9_retrieval_only.json
make reranker   # hybrid + reranker, full judged     → eval/results/hybrid_rerank_k9.json

---

Roadmap

• Document ingestion pipeline (25 documents, 6 sources, ~762K chars)
• Text preprocessing and chunking (500-char overlapping windows)
• Embedding and ChromaDB indexing (all-MiniLM-L6-v2)
• RAG pipeline (retrieval + Groq LLM)
• Streamlit chat interface
• Evaluation framework: retrieval (hit rate, MRR, evidence recall) + answer (fact recall, similarity, NLI) + sampled LLM judge; 40 calls/run vs 1,300+ for Ragas
• Baseline benchmark (30 questions, dense-only): Hit Rate 0.90 / MRR 0.70 at k=7; weak MRR reveals ranking failure — dense embeddings treat year-agnostic chunks as equally relevant
• Temporal disambiguation failure diagnosed: year-specific Budget queries retrieve wrong-year chunks
• Retrieval upgrade — hybrid BM25 + dense + RRF + cross-encoder reranker: Hit Rate 0.97 / MRR 0.80 / Evidence Recall 0.74 vs dense baseline
• k=9 selected via retrieval-only k-sweep (0 Groq calls): top_k controls post-rerank context size; evidence recall maximised at k=9 (0.7389) with flat Hit Rate/MRR
• Reranker treated as optional precision layer (included in production pipeline as robustness measure)
• Eval --mode flag: baseline / hybrid / hybrid_rerank; --retrieval-only for 0-quota k-sweeps
• Hybrid-only ablation: reranker adds no measurable retrieval gain on this corpus at k=9 (Hit Rate / Evidence Recall identical; MRR Δ = 0.0008)
• Dense-side metadata filtering: source/year filter constrains Chroma dense retrieval for targeted queries — Hit Rate 0.9667→1.0000, MRR +2.4pp, Evidence Recall +3.3pp; Q22 partial improvement (hit 0→1, evidence 0→0.5)
• Agentic / multi-modal retrieval: router choosing between policy RAG and structured HDB resale data queries
• FastAPI backend + Docker
• AWS EC2 deployment