e-hekim — Turkish Medical Semantic Search + RAG

An end-to-end system for semantic search and retrieval-augmented generation over a vector database built from health articles published by 14 Turkish hospitals.

Two modes are selectable from a single interface:

Mode	API key	What it does
Semantic search	not required	Vectorizes the question and returns chunks ranked by cosine similarity. This mode alone is enough to evaluate the project without any credentials.
RAG	the user's own key	Passes the chunks that clear the threshold to an LLM and produces a Turkish answer with [1], [2] citations.

The corpus and the user interface are Turkish, because the source articles are Turkish; this document is in English.

---

Two independent refusal layers

Preventing hallucination needs more than a similarity cut-off, so the system refuses in two distinct places and reports which one fired via refusal_reason.

Layer 1 — retrieval gate (below_threshold). If the best-matching chunk scores below the cosine threshold, our own code emits the refusal and the LLM is never called at all. No prompt can talk the system out of this, because no prompt is ever sent.

Bu sorunun cevabı belgelerimde bulunmamaktadır. ("The answer to this question is not found in my documents.")

Layer 2 — model gate (model_insufficient_context). A similarity score cannot tell whether a passage actually answers a question, only that it is on the same topic. "What is Hodgkin lymphoma?" and "What is the five-year survival rate in Hodgkin lymphoma?" retrieve the same chunk with a high score, yet only the first is answerable from it. The model is therefore instructed — as the opening principle of its system prompt — that it has no knowledge of its own for this task, and must decline rather than fill the gap from its pretrained knowledge:

Bu bilgiyi bilmiyorum; bu konuda size yardımcı olamıyorum. ("I do not know this information; I cannot help you with this.")

Partial answers, hedges such as "the documents do not say, but generally…", and adding even a single detail absent from the passages are all forbidden. Measured behaviour on the live system:

Question	Best similarity	Outcome
"Bitcoin bugün kaç dolar?"	0.4298	Layer 1 — LLM never invoked
"Hodgkin lenfomada 5 yıllık sağkalım oranı yüzde kaç?"	0.6153	Layer 2 — passages passed, model declined
"Eritrositler nerede üretilir ve nerede yıkılır?"	0.5931	Answered, with citation

---

Technology stack

Layer	Choice	Note
Vector database	ChromaDB 1.5 (PersistentClient)	Collection created with hnsw:space=cosine; distance = 1 − cosine.
Embeddings	magibu/embeddingmagibu-200m	768 dimensions, 8,192-token context, L2-normalized output.
Backend	FastAPI + Uvicorn	Binds to 127.0.0.1 only.
Frontend	Dependency-free HTML/CSS/JS	Strict CSP; no inline script or style.
LLM access	OpenAI SDK	Every provider speaks the OpenAI wire format; only base_url changes.
Data	umutertugrul/turkish-hospital-medical-articles	CC BY 4.0, ~25K articles, 14 hospitals.

Supported models — DeepSeek (direct) and OpenRouter (multi-provider):

• deepseek-v4-flash (default, thinking enabled), deepseek-v4-pro
• Via OpenRouter: anthropic/claude-haiku-4.5, openai/gpt-4.1-mini, google/gemini-2.5-flash, meta-llama/llama-3.3-70b-instruct

---

Quick start

Requirements: Python ≥ 3.11 and uv. A GPU is optional — everything runs on CPU, only the initial indexing is slower.

git clone https://huggingface.co/datasets/ErenYanic/e-hekim && cd e-hekim
uv venv && uv pip install -e ".[dev]"

cp .env.example .env          # add HUGGINGFACE_TOKEN (the source dataset is gated)

uv run python scripts/ingest.py       # ~5 min on GPU — 1,000 articles to 2,714 chunks
uv run python scripts/benchmark.py    # threshold analysis (optional, writes a report)
uv run python -m ehekim.api           # http://127.0.0.1:8000

Open http://127.0.0.1:8000. Semantic search works immediately. For RAG, paste your own API key into the field in the interface.

Tests: uv run pytest -q (104 tests).

Note: .env is used only by the offline scripts (downloading the source data and uploading to the Hub). The web application never reads an LLM provider key from the environment under any circumstances.

---

1. Article selection and chunking

Selection. 24,612 raw articles, then cleaning (empty bodies, texts shorter than 400 characters, non-http URLs, cookie/KVKK boilerplate, duplicate URLs and byte-identical bodies), leaves 20,549 eligible articles, from which 1,000 are selected.

Rather than mirroring the raw distribution, selection is balanced across the 14 hospitals (71–72 articles per source). In the raw data Acıbadem (6,071) and Memorial (5,264) alone make up half the eligible pool; proportional sampling would have handed half the index to two institutions' house style and topic choices. An equal quota buys wider medical coverage for the same 1,000 documents, which is what makes both the positive and the negative benchmark questions meaningful. Selection is deterministic under seed=42.

Chunking strategy: paragraph-aware, token-bounded, with overlap (hybrid).

• Target 512 tokens, 64-token overlap, 32-token minimum.
• Paragraph integrity comes first: whole paragraphs are packed greedily until the token budget is exhausted.
• A paragraph that overflows the budget is split into sentences (Turkish abbreviations such as Dr., vb., mg. and initials such as M. Ali are not treated as sentence ends).
• If a single sentence still overflows, it is split on a token window as a last resort.

Why this strategy? The corpus is hospital patient-education prose: short titled sections ("Belirtileri nelerdir?", "Nasıl tedavi edilir?"). Paragraph boundaries are genuine semantic boundaries, and blind N-token splitting routinely severs a symptom list from the condition it belongs to. But paragraph lengths are wildly uneven — a one-line introduction next to a 900-token procedure description — so splitting on \n\n alone yields chunks that are both too small to stand alone and too large to be precise. The hybrid approach avoids both failure modes.

A corpus-specific detail: only 35% of the articles contain blank lines (\n\n); the rest separate paragraphs with a single \n (about 44 line breaks per article on average). The chunker therefore treats any run of newlines as a paragraph boundary. Had it looked for \n\n only, two thirds of the corpus would have been processed as one enormous paragraph.

Result: 1,000 articles produce 2,714 chunks (2.71 per article). Tokens: mean 420, median 477, p95 534, max 586.

2. Vector database schema

data/ehekim_chunks.parquet — the required delivery schema plus auxiliary metadata:

Column	Type	Description
url	string	Source link of the article the chunk belongs to
chunk_text	string	The chunked text
chunk_vector	list<float32>[768]	Embedding vector (L2-normalized)
chunk_id	string	{parent_id}-{index}
parent_id	string	Article identifier (first 16 hex of the URL's SHA-1), the parent-child link
title	string	Article title
__source	string	Source hospital (one of 14)
chunk_index	int	Position within the article
token_count	int	Token count of the chunk

The same data is stored in ChromaDB in the ehekim_chunks collection in cosine space.

3. Embedding model

magibu/embeddingmagibu-200m — 768 dimensions, 8,192-token context, ~200M parameters.

Why it was chosen:

• Turkish-focused. Adapted from a multilingual teacher through tokenizer surgery and offline distillation; its TR-MTEB average of 69.5 and STSbTR Spearman of 0.798 put it close to ytu-ce-cosmos/turkish-e5-large at a substantially smaller size.
• Long context. 8,192 tokens is far more than 512-token chunks need, so moving to larger chunks later would not force a change of model.
• Size/quality balance. 768 dimensions give 2,714 × 768 float32 ≈ 8 MB, and the whole corpus vectorizes in about 4.5 minutes on a laptop GPU (GTX 1650).
• L2-normalized output. The dot product equals cosine similarity directly, so Chroma's cosine distance is exactly 1 − similarity.

⚠️ The model is asymmetric. Queries and documents must be encoded with different prefixes: task: search result | query: and title: <title> | text: . The wrong prefix silently depresses similarities and invalidates the threshold calibration. All encoding therefore goes through a single class (ehekim.embedding.Embedder), and a test asserts that the string we build is byte-identical to the model's own registered prompts. The article title is written into the document prefix with its real value.

4. Evaluation set (30 questions)

The repository ships two viewable tables, selectable in the dataset viewer:

Config	Split	Rows	Contents
chunks (default)	train	2,714	The vector database: url, chunk_text, chunk_vector + metadata
benchmark	test	30	The evaluation set with its measured outcomes

from datasets import load_dataset

chunks = load_dataset("ErenYanic/e-hekim", "chunks",    split="train")
tests  = load_dataset("ErenYanic/e-hekim", "benchmark", split="test")

The benchmark table holds 20 positive questions, each written by reading an actual indexed chunk and paired with the URL of the article that answers it, and 10 negative questions whose answers are certainly absent from the corpus (software, sport, finance, history, space, automotive, veterinary medicine). Columns:

Column	Description
id, label	P01–P20 / N01–N10; positive or negative
question	The question put to the system
topic, expected_answer, expected_url	Ground truth for positives
rationale	Why a negative is out of scope
best_similarity	Highest cosine similarity actually retrieved
expected_source_rank	Rank at which the expected article was retrieved
top_match_title, top_match_url	What the retriever returned first
system_decision, expected_decision, correct	Answer/refuse outcome at the 0.53 threshold

Result: 30/30 correct — all 20 positives answered, all 10 negatives refused.

5. Threshold analysis

scripts/benchmark.py runs that 30-question set through the real retrieval path and sweeps the threshold from 0.20 to 0.90, regenerating both the report and the benchmark table above. Full report: data/threshold_report.md.

Separation is decisive:

Group	Mean	Min	Max
Positive (20)	0.7376	0.5819	0.8784
Negative (10)	0.2749	0.1615	0.4777

There is a 0.1042 gap between the lowest positive and the highest negative, so every threshold in [0.50, 0.58] separates the two sets perfectly (F1 = 1.000, accuracy = 1.000, zero false answers on negatives).

Chosen threshold: 0.53 — the midpoint of that plateau. Picking either edge would leave the system on a cliff: at 0.50 the highest negative (the Bitcoin question, 0.4777) is only 0.02 away, and at 0.58 the lowest positive (pharyngeal cancer, 0.5819) is a mere 0.002 away. The midpoint maximizes the margin against both failure modes.

Threshold	Positives answered	False answers on negatives	F1
0.30	20/20	2/10	0.952
0.45	20/20	1/10	0.976
0.53	20/20	0/10	1.000
0.65	17/20	0/10	0.919
0.80	5/20	0/10	0.333

Source recall: the expected article appears in the top 5 for 20/20 questions, and ranks first for 15/20.

An honest caveat. This measurement is at the article (URL) level. Retrieving the right article does not guarantee retrieving the chunk that carries the answer, and the distinction bites in practice, which is what motivated the next section.

6. Parent-context expansion

Chunking necessarily cuts articles at arbitrary points, and the highest-scoring chunk is not always the one holding the answer sentence. Observed case: for "Eritrositler nerede üretilir ve nerede yıkılır?", chunk 1 of the RBC article scores 0.5931 (it discusses low counts) while chunk 0 — which states verbatim that erythrocytes are produced in red bone marrow and broken down in the spleen — scores 0.5176 and falls below the threshold. Given only the passing chunk, the model correctly refused a question the corpus genuinely answers.

So once the threshold gate has decided the query is in scope, each passing chunk brings its immediate siblings (chunk_index ± 1, via parent_id) along as context.

• The gate is not weakened: expansion happens strictly after it and only around chunks that already cleared it, so it can never turn an out-of-scope question into an answered one.
• The cosine values shown in the interface remain the real, unexpanded scores.
• The numbering given to the model and the list rendered in the interface are identical, and siblings added purely for context are marked — ("komşu bölüm"), so a [1] citation always points at the [1] the user can see.

7. Security

Because the user types an API key into the browser, credential handling is the central design constraint of the project:

• The server never stores a key. It arrives in the X-Provider-Key header (not the URL — Uvicorn's access log records paths, so a key in a query string would leak into our own logs), is used for exactly one call, and then goes out of scope.
• No persistence in the browser. No localStorage, sessionStorage, cookie or URL is used; the key lives only in the input value and for the duration of a single fetch. Reloading the page discards it.
• Structural validation. Before a key is placed in an Authorization header it is checked to be a single line of printable ASCII, which stops CRLF header injection at the door.
• Log and error sanitization. Providers echo the submitted key back in 401 bodies (DeepSeek does). Every log record passes through a scrubbing filter, and every error relayed to the client passes through the same scrubber, which replaces credential-shaped substrings with [REDACTED].
• Response models cannot carry a key. No Pydantic model has such a field, and tests assert the key never appears in a response body.
• No CORS, bound to 127.0.0.1, strict CSP (default-src 'self', no inline script or style), nosniff, frame-ancestors 'none', Cache-Control: no-store on API responses, and a 64 KB request-body cap.
• Prompt injection. Retrieved text is fenced inside a <belgeler> element and declared untrusted in the system prompt; the model is told to execute no instruction found inside it.
• Upload protection. scripts/push_to_hub.py works from an explicit allow-list, then applies a deny-list and a secret scan; on any finding it aborts before sending a single byte, and after uploading it re-lists the remote repository to verify both that nothing sensitive leaked and that every expected file arrived.

8. LLM configuration

client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=...,
    reasoning_effort="medium",                     # as specified in the brief
    extra_body={"thinking": {"type": "enabled"}},  # thinking enabled
)

DeepSeek's documentation defines low | high | max for reasoning_effort and, for compatibility, maps medium onto high. The requested value (medium) is sent verbatim; this provider-side mapping is recorded here rather than silently worked around.

OpenRouter models are called with plain completions (no thinking parameters).

9. Project structure

src/ehekim/
  config.py       Settings, prompts, threshold and refusal constants (holds no secrets)
  chunking.py     Paragraph-aware, token-bounded, overlapping chunker
  corpus.py       Article cleaning, balanced selection, chunk records
  embedding.py    SentenceTransformer wrapper with the asymmetric prompts
  vectorstore.py  ChromaDB (cosine) plus sibling-chunk access
  retrieval.py    Threshold gate, context expansion, RAG prompt, refusal detection
  llm.py          Provider catalogue, OpenAI SDK call, error normalization
  security.py     Key validation and credential scrubbing
  api.py          FastAPI application, security headers, endpoints
scripts/          ingest.py, benchmark.py, push_to_hub.py
frontend/         index.html, app.js, styles.css  (no dependencies)
tests/            104 tests — security, chunking, selection, retrieval, HTTP
data/             benchmark_questions.json, threshold_report.md,
                  benchmark_results.json, ingest_manifest.json, *.parquet

10. API

Endpoint	Key	Description
GET /api/health	—	Readiness and chunk count
GET /api/config	—	Threshold/top-k defaults, provider catalogue, refusal messages
POST /api/search	no	Semantic search; chunks with their cosine values
POST /api/ask	X-Provider-Key	RAG; refuses below the threshold without calling the model
GET /api/docs	—	OpenAPI interface

curl -X POST http://127.0.0.1:8000/api/search \
  -H 'Content-Type: application/json' \
  -d '{"query":"Hodgkin lenfomayı ayıran hücre tipi nedir?","top_k":3,"threshold":0.53}'

Licence and disclaimer

The code is MIT. The source data is CC BY 4.0 (umutertugrul/turkish-hospital-medical-articles). This system is for information only; it is not medical diagnosis, treatment or prescribing advice.