axiom-coding-agent-setup 1.0.12 → 1.1.1

package/.agents/skills/project-design/references/ARCHITECTURE.md

@@ -1,641 +1,641 @@
-# AdaptiveRAG — Architecture
-
-A hybrid Adaptive RAG system. Each query is classified at runtime into one of five execution strategies (`no_retrieval`, `vector_only`, `sql_only`, `hybrid`, `clarify`) and dispatched to the right backend(s). Documents flow through a markdown-first ingestion pipeline; the retrieval layer is hybrid (dense + BM25 + cross-encoder rerank); the SQL layer is read-only with defense in depth.
-
----
-
-## Table of Contents
-
-1. [Goals & Non-Goals](#1-goals--non-goals)
-2. [Core Principles](#2-core-principles)
-3. [System Overview](#3-system-overview)
-4. [Tech Stack](#4-tech-stack)
-5. [Project Structure](#5-project-structure)
-6. [Ingestion Pipeline](#6-ingestion-pipeline)
-7. [Chunking Strategy](#7-chunking-strategy)
-8. [Retrieval Layer](#8-retrieval-layer)
-9. [Adaptive Query Router](#9-adaptive-query-router)
-10. [Tool Layer (Read-Only SQL)](#10-tool-layer-read-only-sql)
-11. [Synthesis & Citations](#11-synthesis--citations)
-12. [Caching & Cost Control](#12-caching--cost-control)
-13. [Observability & Cost Tracking](#13-observability--cost-tracking)
-14. [Evaluation Framework](#14-evaluation-framework)
-15. [Configuration](#15-configuration)
-16. [Implementation Status](#16-implementation-status)
-17. [Decision Log](#17-decision-log)
-18. [Future Work](#18-future-work)
-
----
-
-## 1. Goals & Non-Goals
-
-### Goals
-
-- **Adaptive retrieval** — pick the right strategy per-query, not per-file-extension.
-- **Markdown-first ingestion** — convert every input format to markdown so chunks are header-aware.
-- **Hybrid retrieval at the index layer** — dense embeddings + BM25 + reciprocal-rank fusion + cross-encoder reranker.
-- **Grounded answers with citations** — inline `[n]` markers for chunks, `[DB]` for SQL data, parsed back into structured citations for the UI.
-- **Cost-bounded** — content-hash caches for OCR and embeddings; cheap models for routing, frontier models only for synthesis.
-- **Portfolio-presentable** — clean code, working demo, real metrics in Phase 6.
-
-### Non-Goals
-
-- Multi-tenant SaaS with RBAC.
-- Real-time CDC / database mirroring into vectors. (See decision log: never embed structured data.)
-- Distributed task queue. `FastAPI BackgroundTasks` is enough until proven otherwise.
-- A constellation of MCP servers. One optional MCP surface that wraps the same tools is enough.
-- Production monitoring stack (Prometheus / Grafana / Jaeger). Langfuse for traces is enough.
-
----
-
-## 2. Core Principles
-
-1. **If the answer is a sentence, embed it. If the answer is a number, query it.** Free text goes to vectors; structured data stays in SQL.
-2. **Decide adaptively at query time, not at ingest time.** A PDF can contain prose *and* tables; a SQL row can have a free-text comment. Routing has to see the question, not just the file.
-3. **Markdown is the universal intermediate format.** Every parser output normalizes to markdown before chunking.
-4. **Quality of parsing beats quantity of features.** One excellent ingestion path is better than five mediocre ones.
-5. **Measure before optimizing.** No reranker, no advanced chunking, no MCP — until eval scores justify each addition.
-
----
-
-## 3. System Overview
-
-```
-┌──────────────────────────────────────────────────────────────────────┐
-│                         INGESTION (offline)                           │
-│                                                                       │
-│  File ─► FileTypeDetector ─► ParserRouter ─► Markdown ─► Chunker      │
-│                                  │                                    │
-│                                  ├─ Docling          (default)        │
-│                                  ├─ Qwen3-VL         (image / scan)   │
-│                                  └─ passthrough      (.md, .txt)      │
-│                                                                       │
-│  Markdown ─► MarkdownHeaderSplitter ─► Embedder ─► Qdrant             │
-│                                            │                          │
-│                                            ├─ dense  (text-emb-3)     │
-│                                            └─ sparse (BM25 / IDF)     │
-└──────────────────────────────────────────────────────────────────────┘
-
-┌──────────────────────────────────────────────────────────────────────┐
-│                         QUERY-TIME (online)                           │
-│                                                                       │
-│   user query                                                          │
-│       │                                                               │
-│       ▼                                                               │
-│  ┌──────────────────────┐                                             │
-│  │ AdaptiveRouter       │   cheap LLM classifier                      │
-│  │  → strategy + intent │   (gpt-4.1-nano default)                    │
-│  └────────┬─────────────┘                                             │
-│           │                                                           │
-│   ┌───────┼───────────┬─────────────┬────────────────┐                │
-│   ▼       ▼           ▼             ▼                ▼                │
-│ no_retr.  vector    sql_only      hybrid          clarify             │
-│  (LLM)   ↓ Qdrant   ↓ NL→SQL     ↓ both           (ask user)          │
-│          ↓ + rerank ↓ + execute  ↓ + merge                            │
-│           └──────────┴────────────┘                                   │
-│                      │                                                │
-│                      ▼                                                │
-│             GroundedAnswerer (LLM)                                    │
-│                      │                                                │
-│                      ▼                                                │
-│        answer with [n] / [DB] citations                               │
-└──────────────────────────────────────────────────────────────────────┘
-```
-
----
-
-## 4. Tech Stack
-
-| Component | Library | Role |
-|---|---|---|
-| Document parsing | `docling>=2.92` | Born-digital PDFs, DOCX, PPTX, XLSX, HTML, CSV |
-| OCR | `openai>=2.33` against DashScope OpenAI-compat endpoint | Qwen3-VL-Plus for images and scanned PDFs |
-| LLM framework | `langchain>=1.2`, `langchain-core>=1.3` | Message types, prompts, structured output |
-| LLM client | `langchain-openai>=1.2` | Chat synthesis, router, NL→SQL |
-| Embeddings — dense | `langchain-openai` + `text-embedding-3-small` | 1536-dim, with SHA256 disk cache |
-| Embeddings — sparse | `fastembed>=0.4` (`Qdrant/bm25`) | Local BM25 with IDF, no GPU |
-| Vector DB | `qdrant-client>=1.12` + `langchain-qdrant>=1.1` | Hybrid collection (dense + sparse named vectors) |
-| Reranker | `flashrank>=0.2.9` | Pure-ONNX cross-encoder (`ms-marco-MiniLM-L-12-v2`, ~34 MB) |
-| Splitters | `langchain-text-splitters>=1.1` | Header-aware + recursive fallback |
-| PDF inspection | `pypdfium2>=4.30` | Born-digital heuristic + page rendering |
-| Schema validation | `pydantic>=2.9` | Structured router output, structured NL→SQL |
-| SQL | `sqlalchemy>=2.0.36` + `psycopg[binary]>=3.2.3` | Read-only Postgres tool |
-| Retry | `tenacity>=9.x` | Qwen API resilience |
-| UI | `gradio>=6.13` | Tabbed Chat / Ingest / Convert / Admin demo |
-| Tracing | `langfuse>=4.0` | Per-turn spans, token counts, USD cost |
-| Eval | `deepeval>=3.0` | Faithfulness / answer relevancy / contextual relevancy |
-| Env | `python-dotenv>=1.2` | `.env` config loader |
-
-### Explicitly avoided
-
-- `celery`, `redis` — not needed; `FastAPI BackgroundTasks` is sufficient.
-- `transformers`, `torch`, `accelerate` — Qwen is API-based and FlashRank uses ONNX. Keeps the install lean and dodges Python-3.14 native-extension instability.
-- `watchdog` — explicit upload via UI/API is fine.
-- CDC tooling (Debezium et al.) — query DB live via tool, never sync.
-- Prometheus / Grafana — Langfuse covers it for v1.
-
----
-
-## 5. Project Structure
-
-```
-adaptive-rag/
-├── app.py                          Gradio entry point
-├── pyproject.toml
-├── docker-compose.yml              Qdrant + Postgres
-├── .env.example
-├── ARCHITECTURE.md                 (this file)
-├── README.md
-├── PROJECT_PLAN.md                 Phase-by-phase status
-├── scripts/
-│   ├── init_qdrant.py              Create / recreate the Qdrant collection
-│   └── seed_demo_data.py           Seed Postgres with demo e-commerce data
-│
-├── src/
-│   ├── config/
-│   │   └── settings.py             Single source of truth for all tunables
-│   │
-│   ├── core/                       Document → markdown
-│   │   ├── file_detector.py        Format detection + validation
-│   │   ├── docling_parser.py       Docling-backed parser
-│   │   ├── qwen_parser.py          Qwen3-VL OCR with retry + cache
-│   │   ├── parser_router.py        Picks Docling vs Qwen per file
-│   │   └── converter.py            Public conversion API
-│   │
-│   ├── chunking/                   Markdown → header-aware chunks
-│   │   ├── markdown_chunker.py     Header splitter + recursive fallback
-│   │   └── metadata.py             doc_id (SHA256), chunk_uuid (UUID5)
-│   │
-│   ├── indexing/                   Chunks → Qdrant
-│   │   ├── embeddings.py           Dense (cached) + BM25 sparse
-│   │   ├── qdrant_store.py         Hybrid collection + dedup + library
-│   │   └── pipeline.py             Convert → chunk → upsert
-│   │
-│   ├── retrieval/                  Query → ranked chunks
-│   │   ├── hybrid_search.py        HybridRetriever + RetrievalPipeline
-│   │   └── reranker.py             FlashRank ONNX cross-encoder
-│   │
-│   ├── routing/                    Adaptive router + dispatcher
-│   │   ├── strategies.py           Strategy StrEnum + capability sets
-│   │   ├── prompts.py              Router system prompt + few-shots
-│   │   ├── adaptive_router.py      LLM classifier (structured output)
-│   │   └── dispatcher.py           Compose router + retrieval + SQL + synthesis
-│   │
-│   ├── tools/                      External tools the dispatcher can call
-│   │   └── sql_tool.py             Read-only NL→SQL with safety guards
-│   │
-│   ├── synthesis/                  Chunks (+ SQL) → grounded answer
-│   │   └── response.py             GroundedAnswerer + Citation parsing
-│   │
-│   ├── observability/              Tracing + cost tracking
-│   │   ├── langfuse_client.py      Singleton Langfuse + no-op span() ctx mgr
-│   │   └── cost_tracker.py         Pulls daily metrics from Langfuse REST API
-│   │
-│   ├── eval/                       Golden set + accuracy / DeepEval runners
-│   │   ├── golden.jsonl            Tiny smoke golden set (~one row per strategy)
-│   │   ├── run_routing_eval.py     Router-only accuracy gate (CI-friendly)
-│   │   └── run_deepeval.py         DeepEval runner + JSON + HTML report
-│   │
-│   ├── cache/                      Content-hash caches
-│   │   ├── ocr_cache.py            SHA256-keyed disk cache for OCR markdown
-│   │   └── embedding_cache.py      SHA256-keyed disk cache for vectors
-│   │
-│   ├── utils/
-│   │   └── pdf_inspector.py        Born-digital heuristic + page rendering
-│   │
-│   └── ui/                         Gradio interface
-│       ├── main_ui.py              Tab composition
-│       ├── chat_ui.py              Chat tab (calls AdaptiveDispatcher)
-│       ├── ingest_ui.py            Ingest tab (multi-file upload + library)
-│       ├── markdown_converter_ui.py  Convert tab (single-document preview)
-│       └── admin_ui.py             Admin tab — Langfuse cost dashboard
-│
-└── docs/
-    ├── check_postgres.md           DB inspection cheatsheet
-    └── check_qdrant.md             Vector DB inspection cheatsheet
-```
-
----
-
-## 6. Ingestion Pipeline
-
-### Parser routing
-
-```
-file_type ──┐
-            ├── .md / .txt ──────────────────► passthrough
-            │
-            ├── .pdf ──┬─ "born-digital" ───► Docling   (fast, text layer)
-            │         └─ "scanned"  ────────► Qwen3-VL  (vision)
-            │
-            ├── .docx / .pptx / .xlsx / .html / .csv ──► Docling
-            │
-            └── .png / .jpg / .webp ──► Qwen3-VL  (better than Tesseract on layout)
-```
-
-The "born-digital vs scanned" decision for PDFs is a cheap heuristic: render the text layer of the first three pages and treat the file as scanned if the total extracted character count is below a small threshold. Docling has its own internal OCR fallback (EasyOCR / Tesseract); the heuristic lets us skip that path and use Qwen3-VL when accuracy matters.
-
-The user can override this routing per-file with a "Force Qwen3-VL OCR for PDFs" toggle in the Convert tab.
-
-### Qwen3-VL OCR
-
-Calls the DashScope OpenAI-compatible endpoint. The OCR prompt is intentionally deterministic:
-
-> Extract all text from this image into clean GitHub-flavored Markdown. Preserve table structure with pipe syntax. Preserve heading hierarchy. Do not summarize, do not add commentary. If text is illegible, write `[illegible]`.
-
-`tenacity` handles transient API failures with exponential backoff. The result is content-hash cached so re-uploading the same file (or re-rendering the same page from a multi-page PDF) never spends a second API call.
-
-### Content-hash everything
-
-Three things use SHA256 as a primary key:
-
-| Cache | Key | Stored |
-|---|---|---|
-| OCR | `sha256(image_bytes)` | Markdown text on disk |
-| Embeddings | `sha256(model_name + text)` | Raw `float32` vector bytes on disk |
-| Documents | `sha256(file_bytes)[:16]` | `doc_id` for dedup + library listing |
-
-Re-ingesting the same file replaces its prior chunks in Qdrant atomically (delete-by-`doc_id` then upsert).
-
----
-
-## 7. Chunking Strategy
-
-Two-pass:
-
-1. **`MarkdownHeaderTextSplitter`** splits by `#`, `##`, `###`. Headers are kept in the chunk content and the header path is also written to chunk metadata.
-2. **`RecursiveCharacterTextSplitter`** splits any header-section that exceeds `CHUNK_SIZE` (default 1500 chars). Each sub-chunk inherits the parent's header path.
-
-### Per-chunk metadata
-
-```json
-{
-  "doc_id":        "dc7c3912cd0b003d",
-  "source":        "data/policy.pdf",
-  "filename":      "policy.pdf",
-  "header_path":   "Refund Policy > Eligibility",
-  "chunk_index":   7,
-  "total_chunks":  23,
-  "ingested_at":   "2026-05-09T04:52:24+00:00",
-  "parser":        "docling"      // or "qwen3-vl" or "passthrough"
-}
-```
-
-`chunk_uuid` is generated deterministically via `UUID5(doc_id, chunk_index)` so re-upserts are idempotent.
-
-No `access_level` / `department` / `tags` in v1 — those go in only when a feature actually consumes them.
-
----
-
-## 8. Retrieval Layer
-
-### Qdrant collection: hybrid by default
-
-```python
-client.create_collection(
-    collection_name="adaptive_rag",
-    vectors_config={
-        "dense": models.VectorParams(size=1536, distance=models.Distance.COSINE),
-    },
-    sparse_vectors_config={
-        "bm25": models.SparseVectorParams(modifier=models.Modifier.IDF),
-    },
-)
-```
-
-Both vectors are populated for every chunk at ingest time; queries hit both.
-
-### Query flow
-
-```
-query
-  │
-  ├─► dense embedding   (text-embedding-3-small, cached)
-  ├─► sparse encoding   (FastEmbed BM25, IDF on server)
-  │
-  ▼
-Qdrant `query_points` with prefetch:
-  - prefetch dense  (top RETRIEVAL_PREFETCH_K = 25)
-  - prefetch sparse (top RETRIEVAL_PREFETCH_K = 25)
-  - fusion: server-side RRF
-  │
-  ▼
-FlashRank cross-encoder rerank
-  → top RERANK_TOP_K = 5
-  │
-  ▼
-Pass to GroundedAnswerer with header_path context
-```
-
-The fusion is server-side RRF (Qdrant native), not client-side merging — single round-trip per query.
-
-The reranker uses `ms-marco-MiniLM-L-12-v2` by default (~34 MB ONNX). Lazy first-use download to `CACHE_DIR/flashrank/`. If model loading or scoring fails, the pipeline gracefully falls back to the hybrid-fusion order. Alternatives configurable via `RERANKER_MODEL`:
-
-| Model | Size | Notes |
-|---|---|---|
-| `ms-marco-TinyBERT-L-2-v2` | ~4 MB | Fastest |
-| `ms-marco-MiniLM-L-12-v2` | ~34 MB | **Default** — balanced |
-| `ms-marco-MultiBERT-L-12` | ~150 MB | Multilingual |
-| `rank-T5-flan` | ~110 MB | Best quality |
-
-Hybrid search typically yields **+5–15% retrieval recall** over pure dense; reranking adds another **+10–20% context precision** on top. Numbers will be re-validated against the golden set in Phase 6.
-
----
-
-## 9. Adaptive Query Router
-
-The router is what makes this *Adaptive RAG* (per Jeong et al., 2024 — query-complexity-aware strategy selection) rather than static dispatch.
-
-### Strategies
-
-| Strategy | When | Touches |
-|---|---|---|
-| `no_retrieval` | Greeting, chitchat, generic knowledge, math | LLM only |
-| `vector_only` | Conceptual / "what does our doc say about X" | Qdrant + reranker + LLM with `[n]` citations |
-| `sql_only` | Quantitative / "how many" / "top N" / aggregates | NL→SQL → execute → LLM with `[DB]` citation |
-| `hybrid` | Question needs both narrative AND a number | Vector AND SQL → blended answer |
-| `clarify` | Genuinely ambiguous | One focused follow-up question, no retrieval cost |
-
-### How a decision is made
-
-A single LLM call with structured output. The classifier model is intentionally cheap (`gpt-4.1-nano` by default) — classification doesn't need frontier reasoning, and we want this on the hot path of every chat turn.
-
-```python
-class RouterDecision(BaseModel):
-    strategy: Strategy                   # one of the five above
-    reasoning: str                        # one-sentence justification
-    vector_query: str | None              # optional rephrased search query
-    sql_intent: str | None                # NL description for the SQL tool
-    clarification_question: str | None    # only set when strategy == clarify
-```
-
-The router prompt includes:
-1. Strategy descriptions + few-shot examples (anchors classification).
-2. The actual chat history (so follow-ups like "what about last quarter?" can resolve to a real referent instead of always falling to `clarify`).
-3. A one-line summary of available SQL tables — fetched via `inspect()` once at startup. ~80 tokens. Lets the router decide "this is a database question" vs "this is a docs question."
-
-If `SQL_DATABASE_URL` is unset, the prompt is told "no SQL backend" and a sanitizer downgrades any leaked `sql_only`/`hybrid` decision to `vector_only`. The router can never pick a strategy it can't fulfill.
-
-### Dispatch
-
-`AdaptiveDispatcher.answer(query, history)` is the single entry point the chat UI calls. It:
-
-1. Classifies the query.
-2. For `clarify` / `no_retrieval`: short-circuits without touching retrieval or SQL.
-3. For `vector_only` / `hybrid`: runs the retrieval pipeline.
-4. For `sql_only` / `hybrid`: runs the SQL tool. If the tool fails (DB down, query rejected), records a note and continues with whatever else it has.
-5. Calls the appropriate `GroundedAnswerer` method.
-6. Returns an `AdaptiveAnswer` with strategy, decision, citations, executed SQL, and per-stage timings.
-
-Backends are initialized lazily on first use so the app starts fast even when SQL or OpenAI aren't configured.
-
----
-
-## 10. Tool Layer (Read-Only SQL)
-
-### Defense in depth
-
-The SQL tool is *not* an agent. It runs once per turn, with five layers of safety:
-
-1. **Dedicated read-only Postgres role.** `seed_demo_data.py` creates `adaptive_rag_ro` with `SELECT`-only grants. The app connects as that role.
-2. **Statement-level allowlist.** Only `SELECT` and `WITH` (CTE-resolved-to-SELECT) statements are accepted.
-3. **Forbidden-keyword regex.** Catches `INSERT|UPDATE|DELETE|MERGE|TRUNCATE|DROP|ALTER|CREATE|GRANT|REVOKE|COPY|VACUUM|...` even if the role grants would already block them.
-4. **Per-session statement timeout.** Default 5 seconds (`SQL_QUERY_TIMEOUT_SEC`); runaway plans die fast.
-5. **Implicit row cap.** A `LIMIT N` (default 200, `SQL_ROW_LIMIT`) is appended if the SQL doesn't already cap rows.
-
-Plus `SET TRANSACTION READ ONLY` on every connection — belt + suspenders + bungee.
-
-### NL → SQL
-
-A second LLM call (`gpt-4.1-mini` by default, `SQL_MODEL`) translates the router's `sql_intent` into a single `SELECT`. Schema is included in the prompt — generated once via SQLAlchemy `inspect()` at startup, formatted as DDL-style table descriptions with column types, primary keys, and foreign keys.
-
-Structured output forces a `_SqlOutput { sql: str }` schema, so the LLM can't dump prose around the query.
-
-### Demo dataset
-
-`scripts/seed_demo_data.py` builds a deterministic e-commerce schema:
-
-```
-customers   (100 rows)  — id, name, email, country, signup_date
-products    (50 rows)   — id, name, category, price, stock
-orders      (500 rows)  — id, customer_id, status, total, created_at
-order_items (~1300 rows) — id, order_id, product_id, quantity, unit_price
-refunds     (~35 rows)  — id, order_id, amount, reason, created_at
-```
-
-Idempotent (skips if already populated) with `--recreate` to wipe and reseed. The fixed RNG seed gives the same data every run, so demos and golden-set evals are reproducible.
-
-### Optional MCP surface (Phase 7)
-
-The same `SqlTool` and retrieval pipeline can be exposed as an MCP server so Cursor / Claude Desktop can use them without the Gradio UI. This is the *correct* use of MCP — making the same capabilities reusable across LLM clients — not as a microservice framework.
-
----
-
-## 11. Synthesis & Citations
-
-`GroundedAnswerer` has three modes that the dispatcher selects:
-
-| Method | Used for | Context shape |
-|---|---|---|
-| `answer_direct` | `no_retrieval` | Just the user query + chat history |
-| `answer` | `vector_only` | Numbered passage block `[1]..[N]` |
-| `answer_with_sql` | `sql_only` / `hybrid` | Passages **plus** a "Database query results" section with the executed SQL and a markdown-rendered preview of rows |
-
-### Citation contract
-
-The system prompt requires inline brackets for every claim:
-
-- `[1]`, `[2, 3]` — refer to chunk numbers in the passage block.
-- `[DB]` — refers to the SQL results block (only valid when SQL data is present).
-
-The chat layer parses the LLM's output back into structured citations:
-
-```python
-@dataclass
-class Citation:
-    index: int            # 1-based chunk number
-    label: str            # "policy.pdf > Refunds"
-    snippet: str          # short preview
-    source: str | None
-    doc_id: str | None
-    rank: int
-    rerank_score: float | None
-    hybrid_score: float
-```
-
-Only chunks the LLM actually cited end up in the right-hand Sources panel — clean UI by default, with a debug mode that surfaces all retrieved chunks if the LLM cites none. The `cited_db` boolean flag drives a separate "SQL executed" block in the panel.
-
----
-
-## 12. Caching & Cost Control
-
-Three caches on disk, all SHA256-keyed:
-
-| Cache | Key | Rationale |
-|---|---|---|
-| OCR | `sha256(image_bytes)` | Qwen3-VL is paid per-call. Re-ingest of the same image must never re-OCR. |
-| Dense embeddings | `sha256(model_name + text)` | OpenAI is paid per-token. Re-chunking with the same content must not re-embed. |
-| Reranker model | FlashRank's own | First-use download (~34 MB), reused thereafter. |
-
-Default location is `./.cache/`, override with `CACHE_DIR`.
-
-Cost-per-query in the current configuration is dominated by the synthesis LLM (`gpt-4.1-mini`). Routing (`gpt-4.1-nano`) is roughly 1/10th the cost; SQL translation is one extra mini-call only when needed. The live numbers are tracked by Langfuse — see the next section.
-
----
-
-## 13. Observability & Cost Tracking
-
-Tracing is fully optional. When `LANGFUSE_PUBLIC_KEY` / `LANGFUSE_SECRET_KEY` are unset every helper in `src/observability/` is a no-op stub, so the rest of the application code is not branched on "is tracing on?".
-
-### Span tree per chat turn
-
-`AdaptiveDispatcher.answer()` opens one parent span and the helpers open children:
-
-```
-chat.turn (input=query, history_len)
-├── router.classify       (input=query; output={strategy, reasoning, vector_query, sql_intent})
-├── retrieval.hybrid_search (input=query; metadata={fused_count, rerank_ms, prefetch_k, top_k, model})
-├── tool.sql_execute      (input=intent;  output={sql, row_count, columns, rows_preview})
-└── synthesis.{direct|grounded} (input=ctx_summary; output={answer, cited_indices, cited_db})
-```
-
-The LangChain `CallbackHandler` is also passed into every `ChatOpenAI.invoke(...)` (router, NL→SQL, both synthesis modes), so each LLM call appears as a `GENERATION` observation with token counts + the model price → USD cost computed server-side by Langfuse. No price table is duplicated client-side.
-
-`flush_traces()` is called at the end of each turn so spans appear in the dashboard within a second; we don't rely on the background flush thread to survive a Gradio request boundary.
-
-### Cost tracker
-
-`src/observability/cost_tracker.py` is a tiny `httpx`-based reader for Langfuse's `/api/public/metrics/daily` endpoint. It returns a `CostSummary` dataclass with:
-
-- Total traces / observations / tokens / USD cost in the window
-- Per-model breakdown (calls, in/out tokens, cost)
-- Per-day breakdown
-
-The Admin tab renders the same summary as a Gradio markdown view with a 24h / 7d / 30d window selector. CLI:
-
-```bash
-uv run python -m src.observability.cost_tracker --days 7
-```
-
----
-
-## 14. Evaluation Framework
-
-Two complementary scripts. Both consume the same golden set and write reports under `src/eval/reports/`.
-
-### Golden set — `src/eval/golden.jsonl`
-
-The default file ships as a **token-cheap smoke set**: about five rows (one per strategy). Add more rows to `golden.jsonl` anytime you want deeper coverage — there is no required size.
-
-Each row has `id`, `query`, `expected_strategy`, optional `answer_must_contain` / `expected_sql_keywords`, and `notes`.
-
-### Router-only eval — `run_routing_eval.py`
-
-Hits the `AdaptiveRouter` against every example without running retrieval / SQL / synthesis. Reports overall accuracy, per-strategy breakdown, latency, and a confusion matrix. Exits non-zero if accuracy drops below `--threshold` (default `0.85`) — drop into CI to catch router-prompt regressions cheaply.
-
-### Full pipeline + DeepEval — `run_deepeval.py`
-
-For every example whose `expected_strategy` is `vector_only` or `hybrid`, runs the full dispatcher and feeds the (query, answer, retrieved chunks) triple into DeepEval:
-
-- `FaithfulnessMetric` — claims in the answer are grounded in retrieved context
-- `AnswerRelevancyMetric` — the answer addresses the actual question
-- `ContextualRelevancyMetric` — the retrieved chunks were relevant to the query
-
-None of these metrics require a hand-written reference answer — they work with just the `input`, `actual_output`, and `retrieval_context`. Adding `expected_output` later unlocks `ContextualPrecisionMetric` and `ContextualRecallMetric`.
-
-Outputs:
-
-- `src/eval/reports/deepeval_<timestamp>.json` — raw scores per row
-- `src/eval/reports/deepeval_<timestamp>.html` — self-contained summary report
-
----
-
-## 15. Configuration
-
-Every tunable lives in `src/config/settings.py` — a frozen `Settings` dataclass loaded once from `.env` via `python-dotenv`. Some highlights:
-
-| Setting | Default | What |
-|---|---|---|
-| `OPENAI_API_KEY` | (required) | Embeddings, router, SQL gen, synthesis |
-| `QWEN_API_KEY` | (required for OCR) | DashScope endpoint |
-| `QDRANT_URL` | `http://localhost:6333` (or unset → embedded mode) | Vector DB |
-| `QDRANT_COLLECTION` | `adaptive_rag` | Collection name |
-| `DENSE_MODEL` | `text-embedding-3-small` | Must match `DENSE_SIZE` |
-| `DENSE_SIZE` | `1536` | Vector dimensions |
-| `SPARSE_MODEL` | `Qdrant/bm25` | FastEmbed BM25 |
-| `CHUNK_SIZE` / `CHUNK_OVERLAP` | `1500` / `200` | Recursive splitter |
-| `RETRIEVAL_PREFETCH_K` | `25` | Candidates fetched before rerank |
-| `RERANK_TOP_K` | `5` | Final chunks shown to the LLM |
-| `RERANKER_MODEL` | `ms-marco-MiniLM-L-12-v2` | FlashRank model |
-| `LLM_MODEL` | `gpt-4.1-mini` | Synthesis |
-| `LLM_TEMPERATURE` | `0.2` | |
-| `ROUTER_MODEL` | `gpt-4.1-nano` | Cheap classifier |
-| `SQL_MODEL` | `gpt-4.1-mini` | NL→SQL translator |
-| `SQL_DATABASE_URL` | (unset) | Leave unset to disable `sql_only` / `hybrid` strategies |
-| `SQL_QUERY_TIMEOUT_SEC` | `5` | Per-query Postgres timeout |
-| `SQL_ROW_LIMIT` | `200` | Implicit `LIMIT N` injection |
-| `LANGFUSE_PUBLIC_KEY` | (unset) | Set both Langfuse keys to enable tracing — app is a no-op tracer when missing |
-| `LANGFUSE_SECRET_KEY` | (unset) | — |
-| `LANGFUSE_HOST` | `https://cloud.langfuse.com` | Override for the US region or self-hosted |
-| `CACHE_DIR` | `./.cache` | OCR + embedding caches |
-
----
-
-## 16. Implementation Status
-
-| Phase | Status | Description |
-|---|---|---|
-| 0. Setup & infra | ✅ | Repo, deps, Docker compose, settings |
-| 1. Docling baseline | ✅ | Document → markdown for native formats |
-| 2. Parser router + Qwen | ✅ | Born-digital vs scanned heuristic, Qwen3-VL OCR with cache |
-| 3. Chunking + indexing | ✅ | Header-aware chunks, hybrid Qdrant collection, dedup |
-| 4. Retrieval + chat | ✅ | Hybrid search + RRF + FlashRank + grounded answers with `[n]` citations |
-| 5. Adaptive router + SQL | ✅ | Five-strategy router, read-only SQL tool, `[DB]` citations |
-| 6. Eval + tracing + polish | ✅ | Langfuse spans, cost dashboard, routing-accuracy gate, DeepEval runner with HTML reports |
-| 7. Stretch | ⏸️ | C-RAG self-reflection, multi-hop, MCP server, web fallback |
-
-See `PROJECT_PLAN.md` for the full phase-by-phase task list and acceptance criteria.
-
----
-
-## 17. Decision Log
-
-| Decision | Chosen | Rejected | Reason |
-|---|---|---|---|
-| Routing layer | Query-time LLM classifier | Ingest-time extension matching | A PDF can have prose AND tables; routing must see the question |
-| Document parser | Docling | LangChain native loaders, LlamaParse, Marker | Free, local, table-aware, multi-format |
-| OCR | Qwen3-VL-Plus (API) | GLM-OCR (self-hosted), EasyOCR, Tesseract | No GPU, better quality on complex layouts, cheap per-image |
-| Vector DB | Qdrant | pgvector, Weaviate, Chroma | Best hybrid (dense + sparse) support, server-side RRF |
-| Sparse encoder | BM25 via FastEmbed | SPLADE, no sparse | Free, fast, no GPU |
-| Reranker | FlashRank `ms-marco-MiniLM-L-12-v2` | BGE-reranker-v2-m3 (Torch), Cohere Rerank (paid), ColBERT | Pure ONNX via `onnxruntime` (already pulled by `fastembed`); no Torch / Transformers; sidesteps Python-3.14 native-extension instability we hit with `py-rust-stemmers` |
-| Router LLM | `gpt-4.1-nano` | `gpt-4.1-mini`, `claude-haiku` | Cheapest model that classifies reliably; classification doesn't need frontier reasoning |
-| Synthesis LLM | `gpt-4.1-mini` | `gpt-4.1`, `gpt-4o-mini` | Strong enough to follow citation rules, cheap enough for hot path |
-| Tool protocol | Native Python class with explicit dispatch | OpenAI function calling registry, MCP for everything | We're not an agent loop — the router *picks* a strategy, then we run it. Registry abstraction is dead weight. MCP only when external clients consume. |
-| SQL backend | Postgres in Docker bound to host port 5433 | Default 5432, SQLite, Neon | 5433 sidesteps collisions with host-installed Postgres on 5432; users can swap to Neon by changing `SQL_DATABASE_URL` |
-| Read-only enforcement | Dedicated DB role + statement allowlist + keyword regex + statement_timeout + LIMIT injection + READ ONLY transaction | Just one of those | Defense in depth — any one layer might be misconfigured |
-| Task queue | `BackgroundTasks` | Celery + Redis | YAGNI; add when measured queue depth justifies it |
-| File watching | Manual upload | watchdog filesystem watcher | UI-driven flow is enough; watcher is feature creep |
-| DB sync to vectors | Live SQL tool, query at runtime | CDC (Debezium et al.) | Core principle: never embed structured data |
-| Eval | DeepEval + custom routing-accuracy metric | Ragas (async compat issues on Python 3.14), vibes-based | Portfolio projects without metrics look unfinished; DeepEval has no `nest_asyncio` / `sniffio` issues |
-| Tracing | Langfuse | LangSmith, Prometheus + Grafana + Jaeger | LLM-native, model-/framework-agnostic, MIT-licensed core, generous free tier (50k events/month), self-hosting available — sidesteps the LangSmith vendor lock-in concern |
-| Eval reference answers | Skipped for v1 | Hand-written gold answers per row | DeepEval Faithfulness / Answer Relevancy / Contextual Relevancy don't need them; contextual precision and recall do. Add an `expected_output` field to the golden set when the synthesis prompt is stable enough that recall scores are trustworthy. |
-| Cost table source | Langfuse server-side computation | Maintain price table in repo | Model prices change; Langfuse keeps theirs current. We just read totals back via REST. |
-| Tracing default | Disabled (no-op stubs) | Always-on with warnings | Cleaner OSS UX — works without signup; opt-in by setting two env vars |
-
----
-
-## 18. Future Work
-
-### Phase 7 — Stretch
-
-- **C-RAG self-reflection.** Grade retrieved context for relevance; on low scores, re-route or fall back to web search.
-- **Multi-hop retrieval.** When `clarify` would have been picked, escalate to step-by-step iterative search instead of asking the user.
-- **MCP server.** Expose `search_docs` and `query_sql` so Cursor / Claude Desktop can use the same tools.
-- **Web search fallback.** Tavily / Exa when local context is insufficient.
-- **Streaming responses.** Wire LLM streaming through Gradio for snappier UX.
-- **Multi-collection Qdrant.** Split per-domain (`policies`, `finance`, `technical`) once ingest volume justifies the operational cost.
-- **Image-in-markdown OCR.** Extract images from Docling output, OCR them with Qwen, inline the text into the parent markdown.
-
----
-
-**End of architecture document.**
+# AdaptiveRAG — Architecture
+
+A hybrid Adaptive RAG system. Each query is classified at runtime into one of five execution strategies (`no_retrieval`, `vector_only`, `sql_only`, `hybrid`, `clarify`) and dispatched to the right backend(s). Documents flow through a markdown-first ingestion pipeline; the retrieval layer is hybrid (dense + BM25 + cross-encoder rerank); the SQL layer is read-only with defense in depth.
+
+---
+
+## Table of Contents
+
+1. [Goals & Non-Goals](#1-goals--non-goals)
+2. [Core Principles](#2-core-principles)
+3. [System Overview](#3-system-overview)
+4. [Tech Stack](#4-tech-stack)
+5. [Project Structure](#5-project-structure)
+6. [Ingestion Pipeline](#6-ingestion-pipeline)
+7. [Chunking Strategy](#7-chunking-strategy)
+8. [Retrieval Layer](#8-retrieval-layer)
+9. [Adaptive Query Router](#9-adaptive-query-router)
+10. [Tool Layer (Read-Only SQL)](#10-tool-layer-read-only-sql)
+11. [Synthesis & Citations](#11-synthesis--citations)
+12. [Caching & Cost Control](#12-caching--cost-control)
+13. [Observability & Cost Tracking](#13-observability--cost-tracking)
+14. [Evaluation Framework](#14-evaluation-framework)
+15. [Configuration](#15-configuration)
+16. [Implementation Status](#16-implementation-status)
+17. [Decision Log](#17-decision-log)
+18. [Future Work](#18-future-work)
+
+---
+
+## 1. Goals & Non-Goals
+
+### Goals
+
+- **Adaptive retrieval** — pick the right strategy per-query, not per-file-extension.
+- **Markdown-first ingestion** — convert every input format to markdown so chunks are header-aware.
+- **Hybrid retrieval at the index layer** — dense embeddings + BM25 + reciprocal-rank fusion + cross-encoder reranker.
+- **Grounded answers with citations** — inline `[n]` markers for chunks, `[DB]` for SQL data, parsed back into structured citations for the UI.
+- **Cost-bounded** — content-hash caches for OCR and embeddings; cheap models for routing, frontier models only for synthesis.
+- **Portfolio-presentable** — clean code, working demo, real metrics in Phase 6.
+
+### Non-Goals
+
+- Multi-tenant SaaS with RBAC.
+- Real-time CDC / database mirroring into vectors. (See decision log: never embed structured data.)
+- Distributed task queue. `FastAPI BackgroundTasks` is enough until proven otherwise.
+- A constellation of MCP servers. One optional MCP surface that wraps the same tools is enough.
+- Production monitoring stack (Prometheus / Grafana / Jaeger). Langfuse for traces is enough.
+
+---
+
+## 2. Core Principles
+
+1. **If the answer is a sentence, embed it. If the answer is a number, query it.** Free text goes to vectors; structured data stays in SQL.
+2. **Decide adaptively at query time, not at ingest time.** A PDF can contain prose *and* tables; a SQL row can have a free-text comment. Routing has to see the question, not just the file.
+3. **Markdown is the universal intermediate format.** Every parser output normalizes to markdown before chunking.
+4. **Quality of parsing beats quantity of features.** One excellent ingestion path is better than five mediocre ones.
+5. **Measure before optimizing.** No reranker, no advanced chunking, no MCP — until eval scores justify each addition.
+
+---
+
+## 3. System Overview
+
+```
+┌──────────────────────────────────────────────────────────────────────┐
+│                         INGESTION (offline)                           │
+│                                                                       │
+│  File ─► FileTypeDetector ─► ParserRouter ─► Markdown ─► Chunker      │
+│                                  │                                    │
+│                                  ├─ Docling          (default)        │
+│                                  ├─ Qwen3-VL         (image / scan)   │
+│                                  └─ passthrough      (.md, .txt)      │
+│                                                                       │
+│  Markdown ─► MarkdownHeaderSplitter ─► Embedder ─► Qdrant             │
+│                                            │                          │
+│                                            ├─ dense  (text-emb-3)     │
+│                                            └─ sparse (BM25 / IDF)     │
+└──────────────────────────────────────────────────────────────────────┘
+
+┌──────────────────────────────────────────────────────────────────────┐
+│                         QUERY-TIME (online)                           │
+│                                                                       │
+│   user query                                                          │
+│       │                                                               │
+│       ▼                                                               │
+│  ┌──────────────────────┐                                             │
+│  │ AdaptiveRouter       │   cheap LLM classifier                      │
+│  │  → strategy + intent │   (gpt-4.1-nano default)                    │
+│  └────────┬─────────────┘                                             │
+│           │                                                           │
+│   ┌───────┼───────────┬─────────────┬────────────────┐                │
+│   ▼       ▼           ▼             ▼                ▼                │
+│ no_retr.  vector    sql_only      hybrid          clarify             │
+│  (LLM)   ↓ Qdrant   ↓ NL→SQL     ↓ both           (ask user)          │
+│          ↓ + rerank ↓ + execute  ↓ + merge                            │
+│           └──────────┴────────────┘                                   │
+│                      │                                                │
+│                      ▼                                                │
+│             GroundedAnswerer (LLM)                                    │
+│                      │                                                │
+│                      ▼                                                │
+│        answer with [n] / [DB] citations                               │
+└──────────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## 4. Tech Stack
+
+| Component | Library | Role |
+|---|---|---|
+| Document parsing | `docling>=2.92` | Born-digital PDFs, DOCX, PPTX, XLSX, HTML, CSV |
+| OCR | `openai>=2.33` against DashScope OpenAI-compat endpoint | Qwen3-VL-Plus for images and scanned PDFs |
+| LLM framework | `langchain>=1.2`, `langchain-core>=1.3` | Message types, prompts, structured output |
+| LLM client | `langchain-openai>=1.2` | Chat synthesis, router, NL→SQL |
+| Embeddings — dense | `langchain-openai` + `text-embedding-3-small` | 1536-dim, with SHA256 disk cache |
+| Embeddings — sparse | `fastembed>=0.4` (`Qdrant/bm25`) | Local BM25 with IDF, no GPU |
+| Vector DB | `qdrant-client>=1.12` + `langchain-qdrant>=1.1` | Hybrid collection (dense + sparse named vectors) |
+| Reranker | `flashrank>=0.2.9` | Pure-ONNX cross-encoder (`ms-marco-MiniLM-L-12-v2`, ~34 MB) |
+| Splitters | `langchain-text-splitters>=1.1` | Header-aware + recursive fallback |
+| PDF inspection | `pypdfium2>=4.30` | Born-digital heuristic + page rendering |
+| Schema validation | `pydantic>=2.9` | Structured router output, structured NL→SQL |
+| SQL | `sqlalchemy>=2.0.36` + `psycopg[binary]>=3.2.3` | Read-only Postgres tool |
+| Retry | `tenacity>=9.x` | Qwen API resilience |
+| UI | `gradio>=6.13` | Tabbed Chat / Ingest / Convert / Admin demo |
+| Tracing | `langfuse>=4.0` | Per-turn spans, token counts, USD cost |
+| Eval | `deepeval>=3.0` | Faithfulness / answer relevancy / contextual relevancy |
+| Env | `python-dotenv>=1.2` | `.env` config loader |
+
+### Explicitly avoided
+
+- `celery`, `redis` — not needed; `FastAPI BackgroundTasks` is sufficient.
+- `transformers`, `torch`, `accelerate` — Qwen is API-based and FlashRank uses ONNX. Keeps the install lean and dodges Python-3.14 native-extension instability.
+- `watchdog` — explicit upload via UI/API is fine.
+- CDC tooling (Debezium et al.) — query DB live via tool, never sync.
+- Prometheus / Grafana — Langfuse covers it for v1.
+
+---
+
+## 5. Project Structure
+
+```
+adaptive-rag/
+├── app.py                          Gradio entry point
+├── pyproject.toml
+├── docker-compose.yml              Qdrant + Postgres
+├── .env.example
+├── ARCHITECTURE.md                 (this file)
+├── README.md
+├── PROJECT_PLAN.md                 Phase-by-phase status
+├── scripts/
+│   ├── init_qdrant.py              Create / recreate the Qdrant collection
+│   └── seed_demo_data.py           Seed Postgres with demo e-commerce data
+│
+├── src/
+│   ├── config/
+│   │   └── settings.py             Single source of truth for all tunables
+│   │
+│   ├── core/                       Document → markdown
+│   │   ├── file_detector.py        Format detection + validation
+│   │   ├── docling_parser.py       Docling-backed parser
+│   │   ├── qwen_parser.py          Qwen3-VL OCR with retry + cache
+│   │   ├── parser_router.py        Picks Docling vs Qwen per file
+│   │   └── converter.py            Public conversion API
+│   │
+│   ├── chunking/                   Markdown → header-aware chunks
+│   │   ├── markdown_chunker.py     Header splitter + recursive fallback
+│   │   └── metadata.py             doc_id (SHA256), chunk_uuid (UUID5)
+│   │
+│   ├── indexing/                   Chunks → Qdrant
+│   │   ├── embeddings.py           Dense (cached) + BM25 sparse
+│   │   ├── qdrant_store.py         Hybrid collection + dedup + library
+│   │   └── pipeline.py             Convert → chunk → upsert
+│   │
+│   ├── retrieval/                  Query → ranked chunks
+│   │   ├── hybrid_search.py        HybridRetriever + RetrievalPipeline
+│   │   └── reranker.py             FlashRank ONNX cross-encoder
+│   │
+│   ├── routing/                    Adaptive router + dispatcher
+│   │   ├── strategies.py           Strategy StrEnum + capability sets
+│   │   ├── prompts.py              Router system prompt + few-shots
+│   │   ├── adaptive_router.py      LLM classifier (structured output)
+│   │   └── dispatcher.py           Compose router + retrieval + SQL + synthesis
+│   │
+│   ├── tools/                      External tools the dispatcher can call
+│   │   └── sql_tool.py             Read-only NL→SQL with safety guards
+│   │
+│   ├── synthesis/                  Chunks (+ SQL) → grounded answer
+│   │   └── response.py             GroundedAnswerer + Citation parsing
+│   │
+│   ├── observability/              Tracing + cost tracking
+│   │   ├── langfuse_client.py      Singleton Langfuse + no-op span() ctx mgr
+│   │   └── cost_tracker.py         Pulls daily metrics from Langfuse REST API
+│   │
+│   ├── eval/                       Golden set + accuracy / DeepEval runners
+│   │   ├── golden.jsonl            Tiny smoke golden set (~one row per strategy)
+│   │   ├── run_routing_eval.py     Router-only accuracy gate (CI-friendly)
+│   │   └── run_deepeval.py         DeepEval runner + JSON + HTML report
+│   │
+│   ├── cache/                      Content-hash caches
+│   │   ├── ocr_cache.py            SHA256-keyed disk cache for OCR markdown
+│   │   └── embedding_cache.py      SHA256-keyed disk cache for vectors
+│   │
+│   ├── utils/
+│   │   └── pdf_inspector.py        Born-digital heuristic + page rendering
+│   │
+│   └── ui/                         Gradio interface
+│       ├── main_ui.py              Tab composition
+│       ├── chat_ui.py              Chat tab (calls AdaptiveDispatcher)
+│       ├── ingest_ui.py            Ingest tab (multi-file upload + library)
+│       ├── markdown_converter_ui.py  Convert tab (single-document preview)
+│       └── admin_ui.py             Admin tab — Langfuse cost dashboard
+│
+└── docs/
+    ├── check_postgres.md           DB inspection cheatsheet
+    └── check_qdrant.md             Vector DB inspection cheatsheet
+```
+
+---
+
+## 6. Ingestion Pipeline
+
+### Parser routing
+
+```
+file_type ──┐
+            ├── .md / .txt ──────────────────► passthrough
+            │
+            ├── .pdf ──┬─ "born-digital" ───► Docling   (fast, text layer)
+            │         └─ "scanned"  ────────► Qwen3-VL  (vision)
+            │
+            ├── .docx / .pptx / .xlsx / .html / .csv ──► Docling
+            │
+            └── .png / .jpg / .webp ──► Qwen3-VL  (better than Tesseract on layout)
+```
+
+The "born-digital vs scanned" decision for PDFs is a cheap heuristic: render the text layer of the first three pages and treat the file as scanned if the total extracted character count is below a small threshold. Docling has its own internal OCR fallback (EasyOCR / Tesseract); the heuristic lets us skip that path and use Qwen3-VL when accuracy matters.
+
+The user can override this routing per-file with a "Force Qwen3-VL OCR for PDFs" toggle in the Convert tab.
+
+### Qwen3-VL OCR
+
+Calls the DashScope OpenAI-compatible endpoint. The OCR prompt is intentionally deterministic:
+
+> Extract all text from this image into clean GitHub-flavored Markdown. Preserve table structure with pipe syntax. Preserve heading hierarchy. Do not summarize, do not add commentary. If text is illegible, write `[illegible]`.
+
+`tenacity` handles transient API failures with exponential backoff. The result is content-hash cached so re-uploading the same file (or re-rendering the same page from a multi-page PDF) never spends a second API call.
+
+### Content-hash everything
+
+Three things use SHA256 as a primary key:
+
+| Cache | Key | Stored |
+|---|---|---|
+| OCR | `sha256(image_bytes)` | Markdown text on disk |
+| Embeddings | `sha256(model_name + text)` | Raw `float32` vector bytes on disk |
+| Documents | `sha256(file_bytes)[:16]` | `doc_id` for dedup + library listing |
+
+Re-ingesting the same file replaces its prior chunks in Qdrant atomically (delete-by-`doc_id` then upsert).
+
+---
+
+## 7. Chunking Strategy
+
+Two-pass:
+
+1. **`MarkdownHeaderTextSplitter`** splits by `#`, `##`, `###`. Headers are kept in the chunk content and the header path is also written to chunk metadata.
+2. **`RecursiveCharacterTextSplitter`** splits any header-section that exceeds `CHUNK_SIZE` (default 1500 chars). Each sub-chunk inherits the parent's header path.
+
+### Per-chunk metadata
+
+```json
+{
+  "doc_id":        "dc7c3912cd0b003d",
+  "source":        "data/policy.pdf",
+  "filename":      "policy.pdf",
+  "header_path":   "Refund Policy > Eligibility",
+  "chunk_index":   7,
+  "total_chunks":  23,
+  "ingested_at":   "2026-05-09T04:52:24+00:00",
+  "parser":        "docling"      // or "qwen3-vl" or "passthrough"
+}
+```
+
+`chunk_uuid` is generated deterministically via `UUID5(doc_id, chunk_index)` so re-upserts are idempotent.
+
+No `access_level` / `department` / `tags` in v1 — those go in only when a feature actually consumes them.
+
+---
+
+## 8. Retrieval Layer
+
+### Qdrant collection: hybrid by default
+
+```python
+client.create_collection(
+    collection_name="adaptive_rag",
+    vectors_config={
+        "dense": models.VectorParams(size=1536, distance=models.Distance.COSINE),
+    },
+    sparse_vectors_config={
+        "bm25": models.SparseVectorParams(modifier=models.Modifier.IDF),
+    },
+)
+```
+
+Both vectors are populated for every chunk at ingest time; queries hit both.
+
+### Query flow
+
+```
+query
+  │
+  ├─► dense embedding   (text-embedding-3-small, cached)
+  ├─► sparse encoding   (FastEmbed BM25, IDF on server)
+  │
+  ▼
+Qdrant `query_points` with prefetch:
+  - prefetch dense  (top RETRIEVAL_PREFETCH_K = 25)
+  - prefetch sparse (top RETRIEVAL_PREFETCH_K = 25)
+  - fusion: server-side RRF
+  │
+  ▼
+FlashRank cross-encoder rerank
+  → top RERANK_TOP_K = 5
+  │
+  ▼
+Pass to GroundedAnswerer with header_path context
+```
+
+The fusion is server-side RRF (Qdrant native), not client-side merging — single round-trip per query.
+
+The reranker uses `ms-marco-MiniLM-L-12-v2` by default (~34 MB ONNX). Lazy first-use download to `CACHE_DIR/flashrank/`. If model loading or scoring fails, the pipeline gracefully falls back to the hybrid-fusion order. Alternatives configurable via `RERANKER_MODEL`:
+
+| Model | Size | Notes |
+|---|---|---|
+| `ms-marco-TinyBERT-L-2-v2` | ~4 MB | Fastest |
+| `ms-marco-MiniLM-L-12-v2` | ~34 MB | **Default** — balanced |
+| `ms-marco-MultiBERT-L-12` | ~150 MB | Multilingual |
+| `rank-T5-flan` | ~110 MB | Best quality |
+
+Hybrid search typically yields **+5–15% retrieval recall** over pure dense; reranking adds another **+10–20% context precision** on top. Numbers will be re-validated against the golden set in Phase 6.
+
+---
+
+## 9. Adaptive Query Router
+
+The router is what makes this *Adaptive RAG* (per Jeong et al., 2024 — query-complexity-aware strategy selection) rather than static dispatch.
+
+### Strategies
+
+| Strategy | When | Touches |
+|---|---|---|
+| `no_retrieval` | Greeting, chitchat, generic knowledge, math | LLM only |
+| `vector_only` | Conceptual / "what does our doc say about X" | Qdrant + reranker + LLM with `[n]` citations |
+| `sql_only` | Quantitative / "how many" / "top N" / aggregates | NL→SQL → execute → LLM with `[DB]` citation |
+| `hybrid` | Question needs both narrative AND a number | Vector AND SQL → blended answer |
+| `clarify` | Genuinely ambiguous | One focused follow-up question, no retrieval cost |
+
+### How a decision is made
+
+A single LLM call with structured output. The classifier model is intentionally cheap (`gpt-4.1-nano` by default) — classification doesn't need frontier reasoning, and we want this on the hot path of every chat turn.
+
+```python
+class RouterDecision(BaseModel):
+    strategy: Strategy                   # one of the five above
+    reasoning: str                        # one-sentence justification
+    vector_query: str | None              # optional rephrased search query
+    sql_intent: str | None                # NL description for the SQL tool
+    clarification_question: str | None    # only set when strategy == clarify
+```
+
+The router prompt includes:
+1. Strategy descriptions + few-shot examples (anchors classification).
+2. The actual chat history (so follow-ups like "what about last quarter?" can resolve to a real referent instead of always falling to `clarify`).
+3. A one-line summary of available SQL tables — fetched via `inspect()` once at startup. ~80 tokens. Lets the router decide "this is a database question" vs "this is a docs question."
+
+If `SQL_DATABASE_URL` is unset, the prompt is told "no SQL backend" and a sanitizer downgrades any leaked `sql_only`/`hybrid` decision to `vector_only`. The router can never pick a strategy it can't fulfill.
+
+### Dispatch
+
+`AdaptiveDispatcher.answer(query, history)` is the single entry point the chat UI calls. It:
+
+1. Classifies the query.
+2. For `clarify` / `no_retrieval`: short-circuits without touching retrieval or SQL.
+3. For `vector_only` / `hybrid`: runs the retrieval pipeline.
+4. For `sql_only` / `hybrid`: runs the SQL tool. If the tool fails (DB down, query rejected), records a note and continues with whatever else it has.
+5. Calls the appropriate `GroundedAnswerer` method.
+6. Returns an `AdaptiveAnswer` with strategy, decision, citations, executed SQL, and per-stage timings.
+
+Backends are initialized lazily on first use so the app starts fast even when SQL or OpenAI aren't configured.
+
+---
+
+## 10. Tool Layer (Read-Only SQL)
+
+### Defense in depth
+
+The SQL tool is *not* an agent. It runs once per turn, with five layers of safety:
+
+1. **Dedicated read-only Postgres role.** `seed_demo_data.py` creates `adaptive_rag_ro` with `SELECT`-only grants. The app connects as that role.
+2. **Statement-level allowlist.** Only `SELECT` and `WITH` (CTE-resolved-to-SELECT) statements are accepted.
+3. **Forbidden-keyword regex.** Catches `INSERT|UPDATE|DELETE|MERGE|TRUNCATE|DROP|ALTER|CREATE|GRANT|REVOKE|COPY|VACUUM|...` even if the role grants would already block them.
+4. **Per-session statement timeout.** Default 5 seconds (`SQL_QUERY_TIMEOUT_SEC`); runaway plans die fast.
+5. **Implicit row cap.** A `LIMIT N` (default 200, `SQL_ROW_LIMIT`) is appended if the SQL doesn't already cap rows.
+
+Plus `SET TRANSACTION READ ONLY` on every connection — belt + suspenders + bungee.
+
+### NL → SQL
+
+A second LLM call (`gpt-4.1-mini` by default, `SQL_MODEL`) translates the router's `sql_intent` into a single `SELECT`. Schema is included in the prompt — generated once via SQLAlchemy `inspect()` at startup, formatted as DDL-style table descriptions with column types, primary keys, and foreign keys.
+
+Structured output forces a `_SqlOutput { sql: str }` schema, so the LLM can't dump prose around the query.
+
+### Demo dataset
+
+`scripts/seed_demo_data.py` builds a deterministic e-commerce schema:
+
+```
+customers   (100 rows)  — id, name, email, country, signup_date
+products    (50 rows)   — id, name, category, price, stock
+orders      (500 rows)  — id, customer_id, status, total, created_at
+order_items (~1300 rows) — id, order_id, product_id, quantity, unit_price
+refunds     (~35 rows)  — id, order_id, amount, reason, created_at
+```
+
+Idempotent (skips if already populated) with `--recreate` to wipe and reseed. The fixed RNG seed gives the same data every run, so demos and golden-set evals are reproducible.
+
+### Optional MCP surface (Phase 7)
+
+The same `SqlTool` and retrieval pipeline can be exposed as an MCP server so Cursor / Claude Desktop can use them without the Gradio UI. This is the *correct* use of MCP — making the same capabilities reusable across LLM clients — not as a microservice framework.
+
+---
+
+## 11. Synthesis & Citations
+
+`GroundedAnswerer` has three modes that the dispatcher selects:
+
+| Method | Used for | Context shape |
+|---|---|---|
+| `answer_direct` | `no_retrieval` | Just the user query + chat history |
+| `answer` | `vector_only` | Numbered passage block `[1]..[N]` |
+| `answer_with_sql` | `sql_only` / `hybrid` | Passages **plus** a "Database query results" section with the executed SQL and a markdown-rendered preview of rows |
+
+### Citation contract
+
+The system prompt requires inline brackets for every claim:
+
+- `[1]`, `[2, 3]` — refer to chunk numbers in the passage block.
+- `[DB]` — refers to the SQL results block (only valid when SQL data is present).
+
+The chat layer parses the LLM's output back into structured citations:
+
+```python
+@dataclass
+class Citation:
+    index: int            # 1-based chunk number
+    label: str            # "policy.pdf > Refunds"
+    snippet: str          # short preview
+    source: str | None
+    doc_id: str | None
+    rank: int
+    rerank_score: float | None
+    hybrid_score: float
+```
+
+Only chunks the LLM actually cited end up in the right-hand Sources panel — clean UI by default, with a debug mode that surfaces all retrieved chunks if the LLM cites none. The `cited_db` boolean flag drives a separate "SQL executed" block in the panel.
+
+---
+
+## 12. Caching & Cost Control
+
+Three caches on disk, all SHA256-keyed:
+
+| Cache | Key | Rationale |
+|---|---|---|
+| OCR | `sha256(image_bytes)` | Qwen3-VL is paid per-call. Re-ingest of the same image must never re-OCR. |
+| Dense embeddings | `sha256(model_name + text)` | OpenAI is paid per-token. Re-chunking with the same content must not re-embed. |
+| Reranker model | FlashRank's own | First-use download (~34 MB), reused thereafter. |
+
+Default location is `./.cache/`, override with `CACHE_DIR`.
+
+Cost-per-query in the current configuration is dominated by the synthesis LLM (`gpt-4.1-mini`). Routing (`gpt-4.1-nano`) is roughly 1/10th the cost; SQL translation is one extra mini-call only when needed. The live numbers are tracked by Langfuse — see the next section.
+
+---
+
+## 13. Observability & Cost Tracking
+
+Tracing is fully optional. When `LANGFUSE_PUBLIC_KEY` / `LANGFUSE_SECRET_KEY` are unset every helper in `src/observability/` is a no-op stub, so the rest of the application code is not branched on "is tracing on?".
+
+### Span tree per chat turn
+
+`AdaptiveDispatcher.answer()` opens one parent span and the helpers open children:
+
+```
+chat.turn (input=query, history_len)
+├── router.classify       (input=query; output={strategy, reasoning, vector_query, sql_intent})
+├── retrieval.hybrid_search (input=query; metadata={fused_count, rerank_ms, prefetch_k, top_k, model})
+├── tool.sql_execute      (input=intent;  output={sql, row_count, columns, rows_preview})
+└── synthesis.{direct|grounded} (input=ctx_summary; output={answer, cited_indices, cited_db})
+```
+
+The LangChain `CallbackHandler` is also passed into every `ChatOpenAI.invoke(...)` (router, NL→SQL, both synthesis modes), so each LLM call appears as a `GENERATION` observation with token counts + the model price → USD cost computed server-side by Langfuse. No price table is duplicated client-side.
+
+`flush_traces()` is called at the end of each turn so spans appear in the dashboard within a second; we don't rely on the background flush thread to survive a Gradio request boundary.
+
+### Cost tracker
+
+`src/observability/cost_tracker.py` is a tiny `httpx`-based reader for Langfuse's `/api/public/metrics/daily` endpoint. It returns a `CostSummary` dataclass with:
+
+- Total traces / observations / tokens / USD cost in the window
+- Per-model breakdown (calls, in/out tokens, cost)
+- Per-day breakdown
+
+The Admin tab renders the same summary as a Gradio markdown view with a 24h / 7d / 30d window selector. CLI:
+
+```bash
+uv run python -m src.observability.cost_tracker --days 7
+```
+
+---
+
+## 14. Evaluation Framework
+
+Two complementary scripts. Both consume the same golden set and write reports under `src/eval/reports/`.
+
+### Golden set — `src/eval/golden.jsonl`
+
+The default file ships as a **token-cheap smoke set**: about five rows (one per strategy). Add more rows to `golden.jsonl` anytime you want deeper coverage — there is no required size.
+
+Each row has `id`, `query`, `expected_strategy`, optional `answer_must_contain` / `expected_sql_keywords`, and `notes`.
+
+### Router-only eval — `run_routing_eval.py`
+
+Hits the `AdaptiveRouter` against every example without running retrieval / SQL / synthesis. Reports overall accuracy, per-strategy breakdown, latency, and a confusion matrix. Exits non-zero if accuracy drops below `--threshold` (default `0.85`) — drop into CI to catch router-prompt regressions cheaply.
+
+### Full pipeline + DeepEval — `run_deepeval.py`
+
+For every example whose `expected_strategy` is `vector_only` or `hybrid`, runs the full dispatcher and feeds the (query, answer, retrieved chunks) triple into DeepEval:
+
+- `FaithfulnessMetric` — claims in the answer are grounded in retrieved context
+- `AnswerRelevancyMetric` — the answer addresses the actual question
+- `ContextualRelevancyMetric` — the retrieved chunks were relevant to the query
+
+None of these metrics require a hand-written reference answer — they work with just the `input`, `actual_output`, and `retrieval_context`. Adding `expected_output` later unlocks `ContextualPrecisionMetric` and `ContextualRecallMetric`.
+
+Outputs:
+
+- `src/eval/reports/deepeval_<timestamp>.json` — raw scores per row
+- `src/eval/reports/deepeval_<timestamp>.html` — self-contained summary report
+
+---
+
+## 15. Configuration
+
+Every tunable lives in `src/config/settings.py` — a frozen `Settings` dataclass loaded once from `.env` via `python-dotenv`. Some highlights:
+
+| Setting | Default | What |
+|---|---|---|
+| `OPENAI_API_KEY` | (required) | Embeddings, router, SQL gen, synthesis |
+| `QWEN_API_KEY` | (required for OCR) | DashScope endpoint |
+| `QDRANT_URL` | `http://localhost:6333` (or unset → embedded mode) | Vector DB |
+| `QDRANT_COLLECTION` | `adaptive_rag` | Collection name |
+| `DENSE_MODEL` | `text-embedding-3-small` | Must match `DENSE_SIZE` |
+| `DENSE_SIZE` | `1536` | Vector dimensions |
+| `SPARSE_MODEL` | `Qdrant/bm25` | FastEmbed BM25 |
+| `CHUNK_SIZE` / `CHUNK_OVERLAP` | `1500` / `200` | Recursive splitter |
+| `RETRIEVAL_PREFETCH_K` | `25` | Candidates fetched before rerank |
+| `RERANK_TOP_K` | `5` | Final chunks shown to the LLM |
+| `RERANKER_MODEL` | `ms-marco-MiniLM-L-12-v2` | FlashRank model |
+| `LLM_MODEL` | `gpt-4.1-mini` | Synthesis |
+| `LLM_TEMPERATURE` | `0.2` | |
+| `ROUTER_MODEL` | `gpt-4.1-nano` | Cheap classifier |
+| `SQL_MODEL` | `gpt-4.1-mini` | NL→SQL translator |
+| `SQL_DATABASE_URL` | (unset) | Leave unset to disable `sql_only` / `hybrid` strategies |
+| `SQL_QUERY_TIMEOUT_SEC` | `5` | Per-query Postgres timeout |
+| `SQL_ROW_LIMIT` | `200` | Implicit `LIMIT N` injection |
+| `LANGFUSE_PUBLIC_KEY` | (unset) | Set both Langfuse keys to enable tracing — app is a no-op tracer when missing |
+| `LANGFUSE_SECRET_KEY` | (unset) | — |
+| `LANGFUSE_HOST` | `https://cloud.langfuse.com` | Override for the US region or self-hosted |
+| `CACHE_DIR` | `./.cache` | OCR + embedding caches |
+
+---
+
+## 16. Implementation Status
+
+| Phase | Status | Description |
+|---|---|---|
+| 0. Setup & infra | ✅ | Repo, deps, Docker compose, settings |
+| 1. Docling baseline | ✅ | Document → markdown for native formats |
+| 2. Parser router + Qwen | ✅ | Born-digital vs scanned heuristic, Qwen3-VL OCR with cache |
+| 3. Chunking + indexing | ✅ | Header-aware chunks, hybrid Qdrant collection, dedup |
+| 4. Retrieval + chat | ✅ | Hybrid search + RRF + FlashRank + grounded answers with `[n]` citations |
+| 5. Adaptive router + SQL | ✅ | Five-strategy router, read-only SQL tool, `[DB]` citations |
+| 6. Eval + tracing + polish | ✅ | Langfuse spans, cost dashboard, routing-accuracy gate, DeepEval runner with HTML reports |
+| 7. Stretch | ⏸️ | C-RAG self-reflection, multi-hop, MCP server, web fallback |
+
+See `PROJECT_PLAN.md` for the full phase-by-phase task list and acceptance criteria.
+
+---
+
+## 17. Decision Log
+
+| Decision | Chosen | Rejected | Reason |
+|---|---|---|---|
+| Routing layer | Query-time LLM classifier | Ingest-time extension matching | A PDF can have prose AND tables; routing must see the question |
+| Document parser | Docling | LangChain native loaders, LlamaParse, Marker | Free, local, table-aware, multi-format |
+| OCR | Qwen3-VL-Plus (API) | GLM-OCR (self-hosted), EasyOCR, Tesseract | No GPU, better quality on complex layouts, cheap per-image |
+| Vector DB | Qdrant | pgvector, Weaviate, Chroma | Best hybrid (dense + sparse) support, server-side RRF |
+| Sparse encoder | BM25 via FastEmbed | SPLADE, no sparse | Free, fast, no GPU |
+| Reranker | FlashRank `ms-marco-MiniLM-L-12-v2` | BGE-reranker-v2-m3 (Torch), Cohere Rerank (paid), ColBERT | Pure ONNX via `onnxruntime` (already pulled by `fastembed`); no Torch / Transformers; sidesteps Python-3.14 native-extension instability we hit with `py-rust-stemmers` |
+| Router LLM | `gpt-4.1-nano` | `gpt-4.1-mini`, `claude-haiku` | Cheapest model that classifies reliably; classification doesn't need frontier reasoning |
+| Synthesis LLM | `gpt-4.1-mini` | `gpt-4.1`, `gpt-4o-mini` | Strong enough to follow citation rules, cheap enough for hot path |
+| Tool protocol | Native Python class with explicit dispatch | OpenAI function calling registry, MCP for everything | We're not an agent loop — the router *picks* a strategy, then we run it. Registry abstraction is dead weight. MCP only when external clients consume. |
+| SQL backend | Postgres in Docker bound to host port 5433 | Default 5432, SQLite, Neon | 5433 sidesteps collisions with host-installed Postgres on 5432; users can swap to Neon by changing `SQL_DATABASE_URL` |
+| Read-only enforcement | Dedicated DB role + statement allowlist + keyword regex + statement_timeout + LIMIT injection + READ ONLY transaction | Just one of those | Defense in depth — any one layer might be misconfigured |
+| Task queue | `BackgroundTasks` | Celery + Redis | YAGNI; add when measured queue depth justifies it |
+| File watching | Manual upload | watchdog filesystem watcher | UI-driven flow is enough; watcher is feature creep |
+| DB sync to vectors | Live SQL tool, query at runtime | CDC (Debezium et al.) | Core principle: never embed structured data |
+| Eval | DeepEval + custom routing-accuracy metric | Ragas (async compat issues on Python 3.14), vibes-based | Portfolio projects without metrics look unfinished; DeepEval has no `nest_asyncio` / `sniffio` issues |
+| Tracing | Langfuse | LangSmith, Prometheus + Grafana + Jaeger | LLM-native, model-/framework-agnostic, MIT-licensed core, generous free tier (50k events/month), self-hosting available — sidesteps the LangSmith vendor lock-in concern |
+| Eval reference answers | Skipped for v1 | Hand-written gold answers per row | DeepEval Faithfulness / Answer Relevancy / Contextual Relevancy don't need them; contextual precision and recall do. Add an `expected_output` field to the golden set when the synthesis prompt is stable enough that recall scores are trustworthy. |
+| Cost table source | Langfuse server-side computation | Maintain price table in repo | Model prices change; Langfuse keeps theirs current. We just read totals back via REST. |
+| Tracing default | Disabled (no-op stubs) | Always-on with warnings | Cleaner OSS UX — works without signup; opt-in by setting two env vars |
+
+---
+
+## 18. Future Work
+
+### Phase 7 — Stretch
+
+- **C-RAG self-reflection.** Grade retrieved context for relevance; on low scores, re-route or fall back to web search.
+- **Multi-hop retrieval.** When `clarify` would have been picked, escalate to step-by-step iterative search instead of asking the user.
+- **MCP server.** Expose `search_docs` and `query_sql` so Cursor / Claude Desktop can use the same tools.
+- **Web search fallback.** Tavily / Exa when local context is insufficient.
+- **Streaming responses.** Wire LLM streaming through Gradio for snappier UX.
+- **Multi-collection Qdrant.** Split per-domain (`policies`, `finance`, `technical`) once ingest volume justifies the operational cost.
+- **Image-in-markdown OCR.** Extract images from Docling output, OCR them with Qwen, inline the text into the parent markdown.
+
+---
+
+**End of architecture document.**