ghimera

Intent-driven web research: discover sources, collect native-language documents, follow evidence gaps, and return a source-cited answer—or an explicit partial result when the evidence or budget is insufficient.

v0.4.6 adds source feeds and original-evidence reuse for new research.

Configured RSS, Atom, sitemap/index and JSON Feed seeds use the same scoped HTTP/Tor fetch, robots, pacing and native evidence path as ordinary documents. Their outgoing links remain separately admitted frontier work. A configured corpus reader can supply retained originals to current assessment, citations and graph expansion without refetching them or treating old judgments as new. Stored source age remains explicitly unknown; cross-run feed refresh and freshness/reranking are still separate requirements. See source feeds, retained evidence and release acceptance.

The scanned-PDF path renders retained originals into bounded pages, calls your explicitly selected private vision and review services, and preserves generated reading provenance through citations, corpus search, graphs and research planning. It does not download models, silently replace native extraction or certify a language's OCR accuracy. Simplified Chinese quality remains unaccepted. See configuration and evidence and release acceptance. Guarded browser collection and the background delivery worker from 0.4.4 remain available. The repository, distribution and import use ghimera. It succeeds the go-spider distribution and chimera implementation. It is not backward-compatible with v0.1.0's spider_core API or spider CLI. Python 3.11+ is required. Some planned browser/document adapters and public-corpus acceptance are still in progress; see the limitations below.

ghimera is an independent library. Supply your own search provider, self-hosted model services, extraction policies and graph profile.

What is implemented

• Goal and intent loops. Collect from configured seeds with GoalLoop, or use ResearchLoop to plan questions, discover sources through an injected search provider, assess gaps, and draft/review an evidence-cited answer.
• MCP discovery. Borrow an explicitly bound MCP session to obtain source leads; default mappings support the web_search envelope. Other tools and result paths are configurable. See MCP leads.
• Configurable discovery routing. Combine explicit open-web and onion lead providers, with per-provider budgets, concurrent fan-out or ordered fallback, and bounded switching when research stalls. Continuation retains spent budgets and each provider's original response. See discovery routing. An onion-index service is not bundled.
• Ahmia index leads. Query an explicitly configured, operator-owned index, or expose its lead payload through an application-owned onion_search MCP tool. Retain index provenance and fetch original documents separately through Tor. See Ahmia integration.
• Human-proxy browser interaction. An explicitly configured CLI can pause while you act in the selected browser, then resume native capture with request-bound terminal input. See interactive collection.
• Local models first. Configured, already-served self-hosted models supply planning, judging, answer generation, review and embeddings. Compatible HTTP interfaces are supported; no external LLM fallback, model weights or model server are included.
• Bounded direct and Tor HTTP. Native v3 onion collection and open-web requests through Tor share the same route policy. Public-network validation, pinned DNS for direct requests, per-hop redirect checks, robots policy, concurrency/rate limits and retries are accounted for before returning data. A failed Tor route never silently falls back to direct access.
• Isolated JavaScript rendering. Patchright runs in a network-isolated Linux worker; the parent fetch boundary handles its permitted HTTP resources, redirects and accounting. Browser binaries are explicitly configured and verified, not downloaded on import. Alternate passive renderers remain planned; separately deployed challenge gateways are explicitly supported.
• Native extraction. Configured HTML fit-Markdown, adaptive locator profiles, language detection, DOCX tables and native PDF text preserve raw bytes beside extracted native-language text. Full PDF/OCR and Marker acceptance remain open.
• Selective visual evidence. Configured infographic admission filters logos before download, bounds raster decoding and runs offline language-routed OCR. Optional local vision interpretation requires a separate image-bound review. Accepted originals and OCR regions remain beside native text; rejected images are not retained or vectorized. See visuals and Pacific OCR.
• Configurable browsing cadence. Nonnegative jitter adds to origin/robots spacing, while 429/configured throttle responses and Retry-After impose shared-origin cooldown without blocking unrelated origins. See browsing cadence.
• Relevance and deduplication. An injected self-hosted encoder scores native text/windows and observed links against pinned reference vectors. Keyword/semantic ranking, encoding budgets, canonical URL handling, SHA-256 and configured near-duplicate grouping retain source-qualified evidence. Similarity is not a calibrated probability or a substitute for a verdict.
• Graphs and audit records. A configurable research graph starts with the intent. Typed harvests, receipts and ledger rows retain configuration, transport, model-call spend, omissions, verdicts, refusals and source hashes. Operational discovery traces are distinct from evidence-supported claims.

Every candidate reaching acceptance receives an accept/reject/hold verdict. A hold gets a second model pass. An intent is marked answered only after the coverage, citations, review and configured confidence checks pass; budget exhaustion is not silently presented as success.

Additions since go-spider 0.2.0

The current working branch also has explicit authorized source sessions, configurable references/citing-source discovery, persistent publisher-locator drift detection with generic recovery and a doctor, and offline model-based PDF layout, tables, OCR and column-aware reading order, durable run journals, and intent-based semantic scoring without a prebuilt reference-vector file, and a configuration-driven Collector facade using actual adapters. These were not included in the old go-spider==0.2.0 wheel and are included in ghimera==0.3.0. See source sessions, references, locator health, and PDF configuration/acceptance, run journals and intent scoring. For the assembled intent-only API and full non-active template, see configured collector and examples/collector.toml. The package also supports explicitly configured ordinary HTML search alongside JSON, and complete research archives retain successful discovery responses with query/fetch bindings. See HTML search and search evidence. Neither mode solves access challenges, and a refusal is not a successful research result. Representative-corpus accuracy and Marker acceptance remain open; passing a controlled document check is not a universal quality claim.

Additions in 0.4.0

This release adds explicitly configured local challenge recovery with FlareSolverr or Byparr (including its Camoufox-backed 2.x wire), private, origin-scoped clearance and guarded content verification; this is not a universal CAPTCHA solver. See challenge recovery. An explicit dedicated Chromium session can also use an application-supplied human assistance port: finish ordinary login or a challenge in that browser, then continue collection in the same session. DOM acquisition remains distinct from HTTP responses throughout extraction, graph evidence, private archives and completed-round resume. Browser egress is operator-managed; verified browser Tor routing and representative-publisher acceptance remain open. A bounded English/Traditional Chinese public-document capture is recorded in publisher evidence. See human-assisted collection. Version 0.4.2 also captures actual PDF/DOCX downloads through the caller's selected BoundPageHumanSession, injected into Collector. An explicit format policy lets scored native links supply previously unknown file URLs; the collector does not need a hard-coded attachment list. Original file bytes enter the same extraction, citation, graph and archive path. Caller-owned tabs remain open, and unknown browser-network usage is not represented as zero. The 0.4.3 inline-body path makes one separate, configured same-origin browser GET; it does not claim navigation bytes were intercepted. Redirected files and verified entitled-publisher/Tor compatibility remain separate acceptance work. See browser downloads and configuration. It also admits hash-pinned owned PDF/DOCX seeds before intent planning through the same document pipeline. See local inputs. Configured semantic extraction now produces native entity/relationship observations as source-local model assertions. Explicit identity-aware planning can turn repeated names, asserted aliases and potentially competing dated claims into source-bound follow-up research questions without merging the original nodes. These additions were not in the immutable ghimera==0.3.0 artifacts. Alias resolution, graph-driven network expansion and organizational accuracy acceptance remain open, not claims of the existing research graph.

Installation

The 0.4.1 source line adds bounded parallel collection and a durable native evidence corpus. The corpus keeps accepted originals, text/OCR/visual provenance and model-bound vectors for later similarity queries. These additions are not in the immutable 0.4.0 wheel; representative Pacific-language quality and the full infrastructure tracker remain explicit acceptance work.

An explicitly composed PersistentCollector automatically appends completed collection/research evidence to that corpus and returns its acknowledgement beside the unchanged original result. Failed handoffs retain completed source work for persistence-only retry; operation-level crash recovery and an unattended service are still separate requirements. See the corpus documentation for usage.

This source line can also use an explicitly bound corpus as a discovery provider, alone or beside configured web/MCP/onion providers. It preserves native query and source observations; returned matches are leads that still go through ordinary collection and citation checks. Cross-language quality and direct cached-source answer reuse remain separate acceptance work.

An explicit delivery outbox can queue complete results before returning and dispatch them concurrently when its configured destination is available. Destination readback precedes acknowledgement and any explicit local-payload pruning. The included local durable destination is not off-host backup; remote adapters and unattended service deployment remain open.

Version 0.4.4 includes a configurable DeliveryWorker and ghimera-delivery command for background retries, JSONL health and optional readback-confirmed outbox cleanup. Previous release artifacts remain immutable. See delivery lifecycle and configuration.

Use a dedicated virtual environment:

python3.11 -m venv .venv
. .venv/bin/activate
python -m pip install 'ghimera==0.4.6'

Install the adapters you intend to configure:

python -m pip install 'ghimera[html,documents,browser]==0.4.6'

The base package contains the typed core, HTTP/Tor transport, research/search and self-hosted model/embedding clients. Extras add pinned HTML, document and Patchright dependencies. They do not install an inference server, browser binary, Tor daemon or PDF/OCR model artifacts. The isolated browser adapter requires Linux, a compatible explicitly supplied Chromium binary and Bubblewrap; other operating systems have not been accepted for that adapter.

Configuration and API

Operational choices are typed, versioned configuration—not Python constants: scope, budgets, endpoints, model identities/revisions, thresholds, private worker directories, browser provenance and direct/Tor policy. Parse once with GhimeraConfig.from_toml(Path(...)); inject the matching collaborators.

The examples are non-active templates. Replace invalid endpoints, contact information, private paths and model identifiers; reference-vector fixtures are not production relevance data. Adapter blocks belong in the main configuration under their named keys, not as unrelated root settings. Supply any model credential separately in memory, only to its authorized exact endpoint; configuration is not credential or destination approval.

Download the starter configuration, or copy it from the repository:

curl --fail --proto '=https' --tlsv1.2 \
  https://raw.githubusercontent.com/ginkorea/ghimera/v0.4.0/examples/chimera.toml \
  --output chimera.toml

An entirely offline smoke example, using explicitly named test doubles:

import asyncio
from pathlib import Path

from ghimera import GhimeraConfig, Goal, GoalLoop, Harvest, Scope
from ghimera.doubles import FakeExtractor, FakeJudge, FakeRoute, KeywordScorer
from ghimera.fetch import FetchLadder


async def main() -> None:
    config = GhimeraConfig.from_toml(Path("chimera.toml"))
    collector = GoalLoop(
        config=config,
        fetcher=FetchLadder((FakeRoute(),)),
        extractor=FakeExtractor(),
        scorer=KeywordScorer(),
        judge=FakeJudge(),
    )
    result = await collector.run(
        Goal(text="ports", seeds=("https://example.org/start",)),
        Scope(allowed_hosts=("example.org",), max_depth=1, content_types=("text/html",)),
    )
    # Reader revalidates retained sources, provenance and receipt accounting.
    restored = Harvest.model_validate_json(result.model_dump_json())
    print(restored.receipt.stop_reason)


asyncio.run(main())

This example makes no network or real model calls and proves no research accuracy. For actual collection bind CurlRoute, configured extraction and scoring adapters, and the self-hosted judge through their ports. For intent-only research, inject those into ResearchLoop alongside GroundedSearch, IntentPlanner, ResearchAnalyst and AnswerReviewer, then call run(ResearchRequest(intent="your research question")). SearxSearch is the implemented search adapter.

Configured intent API

The configured collector assembles real adapters, so applications need not manually wire every port. Unlike the offline smoke above, this needs your configured services and the adapted examples/collector.toml template:

import asyncio
from pathlib import Path

from ghimera import Collector


async def main() -> None:
    collector = Collector.from_toml(Path("collector.toml"), max_config_bytes=100_000)
    result = await collector.run("Your research question")
    print(result.status)


asyncio.run(main())

Use the collector guide to configure actual private models, source routing, extraction, optional graph/journal and separately supplied credentials. This API is included in ghimera==0.3.0, not the old go-spider==0.2.0 wheel. The lower-level APIs remain supported for custom providers and composition.

The package also supplies a configured intent command that retains the full result, original documents and citations in a private, checksum-sealed archive:

python -m ghimera --job /absolute/path/collector-command.toml --max-job-bytes 100000

See the command guide and examples/collector-command.toml. Existing output identities are never overwritten; partial results remain partial. This command is not in the published 0.2.0 wheel.

The following guides cover the existing lower-level wiring:

Scope and limitations

Robots are honored by default. An override requires a recorded, reasoned, exact-host configuration decision; it does not disable login/paywall refusal. Post-release source adds an opt-in local challenge gateway with bounded, exact-origin clearance reuse; see challenge recovery. That addition is not in the published 0.3.0 artifacts and does not guarantee universal CAPTCHA solving or provide an authenticated-site bypass.

The defined-ontology profile provides separately configured semantic verification, quarantined proposals and source-bound coverage gaps that can motivate follow-up discovery. See semantic verification. Model agreement is not corroboration; real organizational extraction quality remains an open acceptance item. An explicit ghimera.semantic-verification/2 recipe additionally separates named-instance/type checks from relationship entailment, direction and dates. Contradictory summaries refuse instead of overriding a failed dimension; the original review recipe remains unchanged. See the non-active dimensioned example. An explicit version-4 batched review profile preserves the complete original proposal while limiting assessments per answer and retaining a separate coverage call. Its bounds and budgets are configured; it does not claim to fix a model that returns no final answer.

The semantic extraction stage can also feed a configured, bounded graph view into follow-up research planning. Queries retain references to the observed entities/relations and omissions remain explicit; see graph-aware planning. Explicit completed-round suspend/resume preserves evidence, acknowledged graph, pending frontier and cumulative budgets across processes, through the library and a versioned resumable command. Alias resolution, fine-grained interrupted-call reconciliation and real organizational-network quality acceptance remain open. Tor routing is a transport capability, not a guarantee of anonymity or authority to access a source.

Collection is not limited to anonymous access. Supply your own authorized cookies or headers through explicitly configured source sessions, with exact origin/path scope and no credential values in receipts. Browser resources use the same parent-owned session boundary. See authorized sessions.

Source/search requests run on the host executing the crawler. Model control is a separate private-service boundary. Your application owns authorization, deployment and storage; no external scheduler or registry is required to import or use the library.

Configurable source expansion follows observed document URLs and discovers candidate citing sources through your search provider. Depth, host policy and budgets live in [references], not in Python. Source hashes and native locators are retained; a citing-source search hit is not proof that a citation exists. See reference expansion.

For durable observations, configure [journal] and pass a unique run_id. The collector persists JSONL events before acknowledgment and seals a completion summary only after receipt reconciliation. Interrupted prefixes remain inspectable without silently refetching sources. See run journals.

Still required for the complete planned spider: the remaining browser adapters, representative publisher/locator acceptance, full PDF/OCR and Marker validation, real reference/cited-by adequacy, real served-model quality/admission and calibrated decision policy and live runtime/egress acceptance. The repository's detailed tracker retains those requirements; this release does not erase them or describe fixture results as real-world model accuracy.

Development

git clone https://github.com/ginkorea/ghimera.git
cd ghimera
uv sync --locked --extra html --extra documents --extra browser --python 3.11
# Explicit browser/isolation paths are required for the complete gate.
export CHIMERA_TEST_BROWSER=/absolute/path/to/compatible/chrome
export CHIMERA_TEST_ISOLATOR=/absolute/path/to/bwrap
bash scripts/gate.sh
uv build --no-sources

The gate checks the actual interpreter/import path, lockfile, Ruff, strict mypy and the entire test suite. Browser tests refuse absent acceptance prerequisites rather than pretending they ran. Evidence records distinguish protocol fixtures, installed-artifact checks, public-corpus acceptance and production activation.

Migration from go-spider

Install ghimera==0.3.0 explicitly; this is a new distribution, not an in-place rename of old PyPI releases. Use from ghimera import Collector, GhimeraConfig, ghimera.* for submodules, and ghimera or python -m ghimera for the command. The legacy root from chimera import Collector, ChimeraConfig and python -m chimera forward to the same implementation. Old nested chimera.* imports must migrate; there is no second implementation or import hook. Existing versioned chimera.* configuration, graph, harvest, journal and result schemas are retained so existing saved evidence does not change identity.

The v0.1.0 spider_core API and old spider command are not supplied; keep go-spider==0.1.0 while migrating those applications. The old cloud-client, VPN-manager and implicit fallback design is not retained. Prototype source remains in Git history.

See CHANGELOG.md for release changes. Josh Gompert maintains the project at ginkorea/ghimera. Licensed under MIT, matching the existing PyPI licence declaration.